This PDF file contains the front matter associated with SPIE Proceedings Volume 13105, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

This study aims to compare and evaluate the performance of four popular deep learning models (CNN, ResNet50, VGG16, and InceptionResNetV2) in the garbage classification task. Waste classification is essential for sustainable development, and automated waste classification systems have become the key to solving the increasing global waste problem. To achieve accurate waste classification, these systems utilize various machine learning algorithms. In recent years, deep learning models have achieved great success in image classification tasks, so they have attracted much attention in garbage classification. In this study, we comprehensively evaluate the performance of CNN, ResNet50, VGG16, and InceptionResNetV2 models in classifying garbage into 12 different categories by comparing metrics such as accuracy, precision, and confusion matrix. The experimental results show that the InceptionResNetV2 model outperforms other models in the garbage classification task, with high precision and recall. This means that using the InceptionResNetV2 model for garbage classification can identify and separate different types of garbage more accurately. Finally, the deep learning model is connected to the CoppeliaSim platform to control the movement of the robotic arm based on the predictive results of the deep learning model.

With the rapid development of UAV technology, the demand for image recognition segmentation acquired from the air is gradually growing. In this paper, we propose an improved U-Net architecture designed to enhance the effectiveness of UAV image segmentation tasks. Firstly, the traditional convolution module is replaced by a depth-separable convolution to reduce the number of parameters and computational cost and to improve the lightweight performance of the model; secondly, an Axial Attention mechanism is introduced to enhance the model's ability to capture long-range dependencies and spatial information; and finally, the Focal Loss is employed to solve the category imbalance problem. We conducted extensive experimental evaluations on the Aeroscapes dataset containing images captured by UAVs from different altitudes, and the experimental results show that the improved U-Net model exhibits significant improvements in the UAV image segmentation task compared to the traditional U-Net model. Our model improves 6.7% in miou and 2.7% in precision by 2.9% and 3.3% in F1 score. Thus, the method proposed in this paper has potential applications for improving the automation and accuracy of aerial image analysis.

Existing view synthesis algorithms often require dozens or even hundreds of photographs taken from different angles to capture more intricate details in the resulting 3D model. While increasing the number of angles can yield better model parameters, it also significantly increases the training cost. Therefore, there is a need to reduce the quantity of input images with different angles while maintaining comparable results. In this paper, we propose SV-HaRF, which aims to reconstruct scenes by using a sparse set of angles (typically less than 100 but not less than 10) to ensure scene model integrity, as the original training sets often consist of over 100 angles. To address this issue, SV-HaRF employs hash encoding on the input sparse image resolutions to enhance the spatial constraint of these viewpoint positions during the scene reconstruction process, thereby improving the ability to restore scene information. SV-HaRF not only demonstrates outstanding performance in synthesizing sparse viewpoints for reducing model angles but also exhibits superior enhancement, compared to existing NeRF and its related variants, when the input image viewpoints remain unchanged. We conducted experiments on the NeRF Blender Synthetic dataset, showcasing significant improvements in scene synthesis achieved by SV-HaRF.

Real-time rendering, crucial in computer graphics, is essential for interactive applications. Real-time realistic rendering poses a significant challenge, traditionally addressed through rasterization. Despite recent strides in graphics card capabilities, real-time ray tracing struggles due to limited samples per pixel, causing noise and aliasing. This paper introduces a heuristic hybrid real-time rendering method. It comprises five stages: deferred rendering stores data in a G-buffer; a heuristic method selects relevant objects for rendering based on cost and relevance criteria; ray tracing handles secondary effects like shadows and reflections; results are blended with rasterization using the G-buffer, and post-processing enhances the final output. This method optimally balances performance and quality, addressing challenges in achieving ideal real-time rendering results, even with advancements like NVIDIA's RTX series. The proposed approach is implemented in a self-developed engine with a novel architecture, showcasing its effectiveness in real-world scenarios.

For high dynamic range image reconstruction in dynamic scenes, this paper proposes a multi-exposure sequence image fusion method based on histogram matching and convolutional neural network. First, image alignment is performed on the reference image based on the histogram matching method. Then, a multi-exposure fusion model based on convolutional neural network is constructed for the aligned multi-exposure sequence images. Finally, the fusion model is trained based on the training data, and the multi-exposure sequence image fusion in dynamic scenes is realized. The experimental results show that the method proposed in this chapter effectively avoids the influence of dynamic scenes on the fusion of multi-exposure sequence images, and at the same time can generate high-quality fused images.

Aiming at a single feature is difficult to accurately describe the complexity and anisotropy of color and texture of colored spun fabric images, and the color and texture of images may affect each other. A colored spun fabric image retrieval method based on decoupled feature is proposed. The color templates and texture templates of the images are extracted and decoupled; at the same time, deep hash coding is used to calculate the similarity; finally, the preliminary retrieval results are reordered using the decoupled feature. In this paper, seven different types of colored spun fabric sample images are used for retrieval, and the Top-10 recall and mAP of this system reach 95.00% and 86.56%, respectively, which improves the recall and mAP of retrieving Top-10 compared to the methods without feature decoupling and those with feature decoupling only for a single feature.

The field of landscape architecture and planning urgently requires the use of visualization and intelligent technologies to optimize traditional design processes. In response to the digital landscape planning needs, this research has designed and implemented a digital landscape planning system based on visualization technology. Through the development of visualization techniques, system framework design, and the implementation of key technologies, we have developed an integrated landscape planning system that combines digital scene modeling, interactive design, and intelligent planning optimization. The research has verified that the system achieves a high level of fidelity in digital scene representation and demonstrates effective human-computer interaction. The system enables the rapid construction of highly faithful digital landscape scenes, assisting designers in creative design and proposal assessment. Additionally, it leverages knowledge rules and successful cases to facilitate interactive intelligent optimization of landscape design proposals. The goal of this research is to develop a digital landscape planning system that integrates digital scene construction and intelligent design optimization. This research provides valuable technological support and practical implementation for the intelligent and digital transformation of landscape planning, contributing to innovation in the field.

In response to the challenges of diverse target scales, numerous similar objects, and target omissions and false positives resulting from object clustering in unmanned aerial vehicle (UAV) aerial image target identification, this paper introduces an improved UAV aerial image target identification algorithm called U-YOLOE . Within the initial YOLOv8 model's backbone network, a dual-route attention mechanism is incorporated to dynamically and sparsely filter out the least relevant features in the feature maps. This enhances the model's ability to capture key information in UAV aerial images, optimizing detector performance. Additionally, U-YOLOE employs WIoUv3 as the bounding box loss function, enhancing convergence speed and regression accuracy. Experimental results on the VisDrone2019 dataset reveal that, in comparison to the baseline model, U-YOLOE provides a 1% gain in Recall/%, a 1.2% gain in mAP0.5/%, and a 0.6% gain in mAP0.5:0.95%. Compared to other mainstream models, it demonstrates better performance in small object detection tasks for UAV aerial images.

To meet the personalized architectural decoration and daily necessities of ceramic products, low cost, and high efficiency to obtain Chinese painting patterns. This paper designs and realizes the pattern generation scheme of Chinese painting based on the diffusion model. The experimental results show that the model has good image generation ability, and the generated image has higher similarity compared with the original image. It can meet the needs of customized daily ceramics and architectural decoration for traditional Chinese painting patterns. The model provides a new idea for the study of image generation.

With the development of China's aging population and the popularization of intelligent technology, the service design applied to the elderly has attracted more and more attention.According to relevant data, the elderly have more leisure time after retirement, especially the young elderly have a higher frequency of street activities. This paper designs a public seat suitable for aging, and realizes the setting of multiple modules through Python sensing control system. The seat is composed of a sedentary reminder module and a seat surface heating module. The seats are equipped with black solar panels for USB、wireless charging and inductive lighting.From the point of view of geriatric ergonomics, the design of the seat is successively designed in the order of sitting - rising - leaning - walking, to improve the functionality and comfort of the seat.

Image segmentation holds significant research significance and practical applications in the fields of computer vision and image processing, particularly in agricultural production, environmental monitoring, and digital botany. However, there is relatively limited research on its application to paper mulberry, prompting us to construct our own semantic segmentation dataset for paper mulberry. This dataset comprises 2576 images, which were meticulously annotated using the eiseg tool. Data augmentation techniques were also employed during the training process. Regarding the model, this study is built upon the DeepLabV3+ architecture, with the introduction of an attention mechanism known as the Squeeze-and-Excitation Block (SE-Block). Furthermore, modifications to the dilation rates of atrous convolutions within the ASPP module were made to achieve a larger receptive field. Experimental results demonstrate a notable improvement in segmentation accuracy, with a gain of 0.8% over the original algorithm. This enhancement is particularly beneficial for precise leaf area calculations.

Diabetic retinopathy is a common complication of diabetes that can lead to vision loss. Accurate detection of diabetic retinopathy is crucial for early intervention and treatment. This paper proposes a diabetic retinopathy detection method based on EfficientNet and optimizes the pooling layer. Experimental results conducted on the dataset provided by APTOS demonstrate that our approach achieves higher accuracy compared to other traditional methods. This study provides strong support for the early diagnosis and treatment of diabetic retinopathy and holds promising clinical applications.

Computer information processing systems face numerous challenges in the context of big data, and optimization design provides an important avenue to enhance system performance in handling large datasets. This paper first analyzes the bottlenecks, inefficiencies, and poor scalability challenges that big data poses to existing systems. Subsequently, it proposes three system optimization design approaches: modular design, pipeline parallel processing, and elastic scaling mechanisms. Specific solutions such as modular segmentation, MapReduce pipeline implementation, and dynamic resource allocation are presented based on these optimization ideas. These optimization strategies and design solutions effectively enhance the system's scalability and processing performance in the face of big data. Therefore, this paper offers important insights and approaches to drive the evolution of computer information processing systems into the era of big data.

Early screening for tumor symptoms through magnetic resonance imaging (MRI) is crucial for the timely diagnosis and treatment of brain tumor patients. To address the challenge of poor prediction accuracy in single deep learning models, a ensemble learning based brain tumor image classification approach with multiple EfficientNet models is proposed in this paper. Multiple EfficientNet models are integrated, and their feature maps are fused through a spatial channel attention module. Experimental results demonstrate a significant improvement in accuracy compared to other single-model prediction methods. The F1-Score for the four classes, Glioma, Meningioma, Pituitary, and Normal, reached 0.95, 0.95, 0.98, and 0.96, respectively. The proposed method provides robust support for the early diagnosis and treatment of brain tumor patients and holds promising prospects for wide clinical applications.

Lace texture, as a manually designed texture image, needs to possess a series of essential aesthetic characteristics, such as periodicity, symmetry, and blank-leaving in artistic design creation. It requires human designers to spend a lot of time and effort, so it is necessary to apply generative models to generate lace images. In image generation tasks, compared to models such as DCGAN, CycleGAN, and ProGAN, although images generated using StyleGAN2 perform well in terms of resolution, texture details, and periodicity, they still perform poorly in terms of symmetry in lace images. To address the above issues, this article proposes an improved model SStyleGAN (Symmetry StyleGAN) based on StyleGAN2. In terms of discriminators, in order to enhance the attention of the proposed model to image symmetry, we have added a symmetry discriminator, that is, SStyleGAN adopts a dual discriminator structure; In terms of generator, in order to improve the similarity of the feature maps on the left and right sides of the lace image, this paper adds a mean square error loss term based on the loss function of StyleGAN2; In terms of noise input, in order to control the symmetry of the lace image at details such as lace edges, the noise of the StyleGAN2 model is modified to a symmetrical structure, so that the noise input itself has symmetry. In addition to the commonly used FID (Fréchet Insertion Distance) in the generative model, we also used the SSIM (Structural Similarity) metric for the evaluation of the experimental results in this article to detect the symmetry of the generated images. The experimental results show that compared to the lace images generated by the StyleGAN2 model, the lace images generated by the model proposed in this paper not only inherit the advantages of the former, but also have symmetry characteristics.

In order to achieve high accuracy identification of photovoltaic panel faults, a photovoltaic panel fault diagnosis method based on deep learning image recognition technology is proposed. Firstly, the residual network was introduced into the convolutional neural network (CNN) to obtain the residual convolution networks (ResNet) required for the study, and then the ResNet was trained by using the photovoltaic panel fault image dataset to determine the model parameters, and then the photovoltaic panel failure research was carried out. Comparative experiments verify the effectiveness of the proposed ResNet-based fault diagnosis model for photovoltaic panels, which shows that the proposed model has strong generalization and robustness.

In real-world scenarios, due to the complexity of the outdoor environment, motion blur, and the geometric features of the signs themselves, the obtained traffic sign images often exhibit severe distortion, which has a negative impact on subsequent feature extraction and classification. Therefore, this article proposes Canny's traffic sign edge extraction algorithm based on bilinear interpolation improvement. Firstly, in order to reduce the actual consumption of image processing and recognition in the later stage, this article adopts the weighted average method to process the grayscale of the original color image; Secondly, in order to reduce the impact of light intensity on image quality, this article adopts a histogram equalization image enhancement method; Then, in order to avoid the impact of scale on subsequent feature extraction and classification, this paper uses bilinear interpolation to normalize the image; Finally, by improving the Canny edge detection algorithm, the problem of limited edge detection for long-distance traffic signs is solved, thereby accurately detecting edges and improving efficiency. These preprocessing steps can effectively improve image quality and improve the accuracy of subsequent feature extraction and classification.

With the growth of electricity demand and the expansion of power grid scale in China, thermal fault identification of electrical equipment in substations plays a crucial role in ensuring the safe and efficient operation of the power grid and preventing accidents. Infrared imaging technology can identify abnormal temperature areas based on the surface temperature distribution of equipment through non-contact, long-distance, and non-stop methods, thereby identifying thermal faults of equipment without the need for power outage. Due to the increasing resolution of infrared images collected by modern infrared detection equipment, traditional image segmentation techniques are difficult to meet real-time requirements. Therefore, this article proposes an image segmentation algorithm that combines genetic algorithm and fuzzy C-means clustering. It improves the selection, crossover, and mutation operators of genetic algorithm and combines them with fuzzy C-means clustering, effectively avoiding the disadvantage of traditional FCM being sensitive to random initial values and easily trapped in local optima. Experiments have shown that this algorithm accelerates convergence speed, improves image segmentation accuracy, improves global search ability, and can effectively improve image segmentation accuracy.

In the realm of computer-aided diagnosis, image segmentation techniques have garnered significant attention over the years. However, medical image data present distinctive characteristics, such as confidentiality, class imbalance, and challenges in acquiring accurate labels. Of certain rare medical conditions, it is arduous to obtain an adequate number of data samples within a limited scope for training models. Moreover, complex medical conditions may lead to diagnostic omissions concerning the locations of minor lesions. Consequently, our primary objective in this article is to address the challenge of enhancing the segmentation accuracy of lesion locations in complex medical conditions when working with limited data. Additionally, we believe that threshold selection and feature choice are particularly crucial for medical image segmentation tasks. However, existing medical image segmentation methods often prioritizes model complexity while overlooking the impact of threshold variations on segmentation performance. In this paper, we propose a multi-feature fusion network with adaptive threshold prediction and introduce a novel paradigm of semi-supervised pseudo-label generation to facilitate supervised learning. The multi-feature fusion module ensures the comprehensive utilization of dimension-sensitive features, while the adaptive threshold predictor enables the model to generate distinct segmentation thresholds based on varying inputs. We employed publicly available medical image datasets, namely Kvasir-SEG, Rim-one r2, and ISIC 2018, to train and validate our model and approach. In comparison to other supervised and semi-supervised methods, the model and the approach we proposed demonstrate superior segmentation performance.

Privacy and security of multimedia data such as images has been an important concern particularly after the extensive use of digital devices in people’s daily life. Therefore, more and more researchers put their attention on investigating new secure image encryption algorithms. However, many of the existing encryption methods contain the issues of computational overhead, substantial time complexity or low resistance to cryptanalytic attacks. To address these issues, we propose a new encryption algorithm strategically constructing a 3D array from pixel array of the plaintext image and utilizing the chaotic logistic map, the AES substitution box, and linear mixing. Performance of the proposed scheme is validated through analyzing the information entropy, histogram, correlation coefficients between adjacent pixels and mean squared errors of ciphertext images. The results exhibited superior randomness and unpredictability of the encrypted image. Besides, the algorithm manifested strong resistance against the differential attacks.

This paper designs a more efficient CNN (Convolutional Neural Network) architecture to extract more spatial features to solve this problem. A Double Channel Convolution(DCC) is designed using two multiple convolution modules to capture more spatial features and a residual module to prevent network performance degradation. The skip connection employs spatial and channel attention mechanisms for dual capturing of associations among global and local features between space and channel to enhance the correlation of features from modalities to modalities as well as the correlation judgment of Region of Interests (ROIs) and boundary information. The effectiveness of the network is verified by the experimental results with open access available dataset BraTS21, which shows that the Dice similarity coefficient (DSC) in the segmented brain tumors in the Enhance Tumor(ET),Tumor Core(TC) and the Whole Tumor(WT) were 0.832, 0.873 and 0.915, which are 1.2%, 1.16% and 2.2% higher than the DSC, JSC and Sensitivity of the U-Net model, respectively. The model size is reduced by 7.6 M. The experimental results show that the ALU-Net model proposed in this paper can achieve better performance and more computational efficient. The model showed good performance in automatic brain tumor segmentation.

For ancient paintings and calligraphies, including thangkas and murals, which are often of large format and cannot be scanned at one time, it is necessary to utilize high-precision CIS image sensors to scan the paintings and calligraphies several times in segments, and then finally get the finished images through image stitching. Among them, fast and accurate stitching of large format images with hundreds of millions of pixels is a major difficulty. Due to the huge size, more computational resources are required, and conventional image stitching algorithms can not directly complete the stitching task in the case of limited performance. Therefore, we propose a high-resolution image stitching algorithm OR-SIFT based on deep learning prediction of overlapping regions, which combines the convolutional neural network with the traditional feature detection method. It uses the neural network to predict the overlapping region as the region of interest for subsequent stitching algorithms. Then, an improved SIFT algorithm is used to extract and describe features in the overlapping region, followed by precise stitching. Additionally, a strategy for continuously stitching multiple high-resolution images is proposed to reduce computational complexity while ensuring accuracy, achieving continuous stitching of multiple high-resolution images.

Electricity is an indispensable part in the development of national economy as well as people's life. Under the fast-paced social development trend, advanced intelligent equipment is widely used in the power system, and the manual power line inspection method is replaced by machines and equipment, which avoids all kinds of difficulties encountered by the staff in the inspection process. However, the power line inspection map captured by the intelligent power line inspection equipment and remotely sent to the dispatching terminal still requires staff to identify and analyze the specific power line inspection situation. In this paper, based on the image recognition technology of deep neural network, we analyze and select Faster R-CNN, a target recognition algorithm with high precision, and build a dedicated power line inspection image recognition model in TensorFlow deep learning platform. The model realizes automatic recognition of different types of power line inspection images transmitted back from remote locations, with high recognition accuracy, low misjudgment rate and omission rate, and can timely and accurately detect and identify hidden dangers in the process of power line inspection. In the power line inspection, the integration of image recognition technology, real-time effective identification of hidden dangers, and save the scheduling of manual work time, improve work efficiency.

The widespread dissemination and sharing of digital images have led to an urgent need for image copyright protection. To address this challenge, this study proposes a comprehensive image copyright protection system based on blockchain and image processing algorithms. Firstly, image features are extracted through binarization processing and the ORB algorithm. The GMS algorithm is then employed for feature selection to enhance the accuracy of feature matching. Subsequently, the InterPlanetary File System (IPFS) is utilized to store images and related data in a decentralized network, ensuring the immutability and traceability of information. Finally, image copyright management throughout its lifecycle, including registration, querying, and transfer, is implemented through Ethereum smart contracts.

An improved particle swarm optimization algorithm was proposed to solve the problem of the optimal layout of fixed-tilt mirrors in tower solar thermal power plants. This article introduced nonlinear inertia weight and asynchronous learning factor. It also provided an example to verify the method. In this paper, the rated annual average thermal power output for a fixed-tilt mirror field reached a certain size. The unit area annual average output of thermal power was taken as the evaluation criterion for optimizing the layout of the mirror field. Firstly, the initial optical efficiency and thermal power of the fixed-tilt mirror field were determined based on the initial layout position and parameters. Then, based on the characteristics of the absorption tower radiating outward, the fixed-tilt mirrors were classified from inside to outside to form a layered structure to optimize the solution dimension. Then, the improved particle swarm algorithm was used to optimize each parameter of the fixed-tilt mirror, to find the global optimal solution and obtain the optimal distribution position and parameters of the fixed-tilt mirror. Finally, the improved algorithm was compared with the standard particle swarm algorithm: from the experimental results, it can be seen that using the improved particle swarm algorithm can find a better hyperparameter for the mirror field layout model. It also proves that the improved algorithm has better convergence speed, robustness, and adaptability, and verifies the correctness and effectiveness of this improved method.

This study utilizes regression analysis to examine the external factors affecting the calibration outcomes of binocular cameras. The experiment employs a binocular camera with fixed focal length and baseline, coupled with a commonly used checkerboard calibration pattern. A series of calibration pattern images were captured and subjected to image processing and parameter computation. Employing factorial experiments, variance analysis was conducted to assess the impact of factors such as black and white spacing between checkerboard patterns, the quantity of calibration images, the number of internal corner points, and the distance between calibration boards on calibration errors. Results indicate that all aforementioned factors influence baseline errors. Findings suggest that optimizing the quantity of calibration images, adjusting black and white spacing in checkerboard patterns, increasing internal corner points, and reducing the distance between calibration boards enhance the precision and performance of binocular cameras in 3D vision applications.

Aiming at the problems of poor brightness effect and poor recognition, a darkroom image enhancement algorithm based on Retinex algorithm is proposed, which replaces Gaussian filter with bilateral filter and adaptively adjusts the saturation of the darkroom image. The algorithm firstly converts the darkroom image from the three-primary color model (RGB) to the hue saturation value model (HSV). The Retinex algorithm is used to obtain the incident component and the reflection component of the darkroom image.The reflection component is sharpened by Laplace algorithm to enhance the brightness of the image and retain the color sense of the original image, and the incident component is processed by bilateral filtering. In this way, the visual display effect of images can be improved. Finally, the adaptive adjustment strategy of correlation coefficient is adopted for the saturation component in HSV space, so that the saturation component can adjust itself and maintain the global color of the image. Compared with histogram equalization (HE) algorithm and single-scale Retinex(SSR) algorithm, this algorithm can effectively solve the problems of color loss and edge information loss in darkroom images.

With the rise of deep learning, the cross collision between artificial intelligence and art, represented by image style transfer, has attracted high attention in the fields of graphic and image technology and art. Based on Convolutional Neural Networks(CNN), this article discusses an image style transfer model. Firstly, define the content loss function and style loss function ,weigh them to calculate the total loss function, incorporate texture features extracted by the Canny operator into the content image; secondly, use the cost function to calculate the values representing content and style for comparison between the initial style and the target style, and then generate a new target image through deformation and combination; finally, train the VGG model to generate the output of the content and style layers of the image, and calculated the corresponding layer parameters applied to the original image for retraining to achieve the minimum loss value. The verification results indicate that adding the Canny operator makes the edges of the style transfer graph more delicate and the contours clearer. This model has a very wide range of applications, such as video processing, photo beautification, social communication, etc. It is currently one of the popular technologies in artificial intelligence.

Due to cost and equipment limitations, steel surface defect images are often of low resolution, significantly impairing the recognizability of key features within the images and thereby reducing the accuracy of automated detection. To address this issue, this paper proposes a classification method for low-resolution steel defect images based on super-resolution knowledge distillation. This method initially employs advanced super-resolution techniques to reconstruct low-resolution images, aiming to restore critical image details and features. Subsequently, knowledge distilled from deep learning models trained on high-resolution images is transferred to models specifically designed for low-resolution images. This approach combines the advantages of super-resolution image reconstruction and the efficiency of knowledge distillation. It not only enhances the quality of low-resolution images but also maintains the lightweight nature of the model, making it suitable for real-time detection scenarios. A series of experiments conducted on publicly available steel defect dataset demonstrate that the method proposed effectively enhances the classification accuracy of low-resolution steel defect images. This achievement is of significant importance in advancing industrial automated defect detection technologies and provides a new research direction in the field of low-resolution image processing.

This paper introduces an effective algorithm for classifying maize leaf diseases using a publicly available image dataset. While the dataset itself, consisting of maize leaf disease images, was developed by Indian researchers for agricultural applications, this study's contribution is the development and optimization of an image classification algorithm tailored for this dataset. The paper details the preprocessing and data augmentation techniques applied, such as size adjustment, normalization, and enhancement to ensure data uniformity and diversity. These methods significantly improve the model's accuracy and generalization capabilities. The study employs a deep learning model based on resnet18, dividing the dataset into training, validation, and test sets in a 6:2:2 ratio. The training process is analyzed through loss and accuracy curves, revealing a steady decrease in loss and increase in accuracy over time. The model achieves 98% accuracy, 95% Precision, 95% Recall, and a 95% F1-score, indicating high effectiveness in classifying maize leaf diseases. A confusion matrix analysis confirms the model's precision across most image types. The research demonstrates the algorithm's strong performance in classifying various types of maize leaf diseases, highlighting its potential to aid farmers in disease identification and treatment, thereby improving agricultural productivity and quality.

Thoracic CT image registration is crucial for thoracic image analysis and related downstream clinical diagnosis and treatment. Due to the large-scale deformation of multiple tissues and organs involved in the respiratory process, achieving high-precision thoracic CT image registration is very challenging. It should be noted that this large-scale deformation occurs in a continuous cycle over time and space. Therefore, incorporating spatio-temporal cycle consistency into the registration can aid in achieving high-precision thoracic CT image registration.However, most existing registration methods either focus only on static registration between two typical respiratory stages (usually extreme exhalation and extreme inhalation), or are based on assumptions of specific respiratory motion models that are difficult to generalize.To this end, this paper proposes a spatio-temporal cycle consistency registration method to embed the continuous cyclic changes of respiratory deformation in time and space. Specifically, we first construct continuous sequential CT image registration in two stages: exhale to inhale and inhale to exhale, to introduce spatio-temporal cyclic consistency of global deformation. In addition, spatio-temporal consistency based on the lung mask is employed to deal with the large-scale deformation in the lung region to achieve better local matching. Finally, based on the spatio-temporal consistency of the deformation of the lung vessel mask, the internal deformation of the lung is further refined to improve the registration accuracy.We conducted experiments and evaluations on the 4D-CT DIR and COPD datasets. Experimental results show that our method outperforms advanced methods by incorporating spatio-temporal cycle consistency.

Brain medical image registration is the foundation of numerous clinical applications in neurosurgery, such as preoperative planning and intraoperative navigation. However, precise registration remains challenging when dealing with large-scale deformations between images to be registered and significant anatomical differences across modalities. Existing methods often incorporate explicit structure information (e.g., anatomical structure segmentation masks) as prior knowledge to assist registration, but obtaining high-precision structure prior knowledge itself is difficult. In addition, most existing methods for large-scale registration use low-resolution advanced semantic relationships as initial driving information for registration, which may mislead the registration process that relies primarily on structure alignment. To address these issues, this paper proposes a novel brain image registration method based on spatial structure inter-correlation. Specifically, we first extract shallow multi-scale anatomical structure information because the core of registration is the alignment of corresponding structures. Compared to deep structures rich in semantic information, shallow structures rich in structure information are more conducive to guiding the registration process. Then, we construct structure space inter-correlation based on the extracted shallow multi-scale anatomical structure information (including structure attention calculation between two modalities). Different scales of structure information contain different levels of structure inter-correlation, and by performing multi-level low-resolution upsampling and fusion, we obtain the final registration result to improve registration accuracy. Finally, starting from dual registration, the deformation from moving images to fixed images should enable fixed images to deform to moving images in the dual direction, thereby improving the robustness of registration. Through comparison with many advanced methods on brain MRI datasets, our proposed method demonstrates superior accuracy in brain images registration.

Objective To explore the interaction principle between visual perception and multimodal image nonlinear editing, and analyze the mode of image nonlinear editing and image coding output from the perspective of multimodal analysis, so as to make the nonlinear editing and image coding output more scientific, rational and harmonious. Methods By the method of multimodal analysis, the principle of visual perception and the principle of image nonlinear editing were used to analyze and discuss the multimodal analysis of image editing, and video editing, audio editing and special effects editing were decomposed and analyzed by visual mode, auditory mode and emotional mode respectively. Conclusion From the view of visual perception and multimodal analysis, a structured method of image processing is proposed to improve the science, rationality and harmony of nonlinear editing of images.

Aiming at the mosaic phenomenon in images, the existing single-feature evaluation indexes of non-referential image quality do not conform to the subjective evaluation of the human eye, and a single-dimensional evaluation standard of edge slope consistency is proposed. For the super-resolution reconstruction image quality evaluation research, the proposed multi-dimensional non-referential image quality evaluation index, can simulate the subjective evaluation of the calculation of image quality real-time scoring, and distinguish between different super-resolution reconstruction algorithm results. Design and analysis of the experimental results, the unidimensional and multi-dimensional image quality evaluation standard proposed in this paper is more in line with the subjective evaluation of the human eye, and at the same time, combined with the database LIVE and QADS compared with the other image evaluation index test, which illustrates the superiority and versatility of the multi-dimensional image quality index.

As a sustainable urban transportation mode, cycling has the advantages of energy saving, environmental protection, relieving congestion, and improving road utilization. However, the vast majority of urban road network data are mainly organized and constructed by motor vehicle services, lacking perfect cycling attribute information, which is difficult to meet the needs of cyclists. Therefore, this paper studies a method to extract the semantic information of cycling lanes from street view images. In this method, the detection algorithm yolov7 is used to construct the cycling traffic sign classification model, extracting semantic information related to bicycles from street view images, and then it is matched with the OSM road network data to generate an updated cycling road network. The experimental results show that this method can effectively identify all kinds of cycling traffic signs in the street view images, and the extracted semantic information has a good matching degree with the OSM road network. Using street view images to extract cycling semantic information in roads is an effective means to enrich the cycling attributes of existing road network data.

Fine-grained image classification aims to accurately categorize subclasses within a particular category. Due to the small inter-class differences and large intra-class variations, fine-grained image classification has been a challenging research topic in the field of computer vision and holds significant research value. Existing neural network-based algorithms suffer from the loss of fine-grained texture details during the training process and the inability to effectively fuse features extracted from different convolution layers of the backbone network. To address these issues, this paper proposes a fine-grained image classification method based on a lightweight feature extraction network with MobileNet v2 as the core, incorporating multi-scale feature fusion and attention mechanism. Considering that high-level and low-level features contain rich semantic and textural information, attention mechanisms are embedded into different scales to capture more diverse feature information. Experimental evaluations conducted on the publicly available fine-grained dataset, A Large Scale Fish Dataset, achieve a classification accuracy of 99.86%. The results demonstrate the superiority of the proposed method in fine-grained object classification.

For the end-to-end Faster R-CNN rotation region detection algorithm, there are problems such as high difficulty in detecting frame coordinate regression and unstable structure. Compared with the detection frame, the pixel mask describes the target area at the pixel level, which is a more accurate annotation method to locate the target, and is often used in semantic segmentation.We further proposed a ship rotation region detection algorithm based on semantic segmentation. We first used an improved Focal Loss function for training the improved U-Net segmentation model integrating the space and channel attention mechanism and feature pyramid network to obtain the ship semantic segmentation mask, and then used the watershed segmentation algorithm to distinguish different instances and did the minimum outer rectangle detection to position ships rotation region after acquiring their area masks. Our method can not only accurately position the ships but also effectively obtain the scale information of them which giving an indication of certain practical value.

As a device for detecting colopathy, white light endoscopy still suffers from missed diagnoses and misdiagnoses. For increasing the diagnostic rate of colopathy sick persons, our work proposes a new diagnoses model for pathological changes based on an improved YOLOv5 algorithm based on the SwinTransformer framework. In this diagnostic model, colopathy is divided into three categories: polyp of the colon, adenoma of the colon, and carcinomas. The SwinStage structure, as well as the channel attention mechanism and spatial attention mechanism, are incorporated to fully extract features of the map and the relationship between feature maps, thereby improving the result of object detection. Experimental outcomes demonstrate that the proposed pathological tissue diagnoses model has a more excellent pathological tissue detection ability, the mAP@0.5 and mAP@0.5:0.95 can attain 0.915 and 0.717.

At present, a large number of light UAV is used in power patrol instead of traditional manual patrol to improve work efficiency, and a large number of detection equipment need to be equipped with lightweight detection algorithm with excellent performance. To solve this problem, an improved YoLov4-tiny algorithm is proposed. The main structural improvement is to add an improved SPP structure to the output layer of the backbone network. Because of the limited sample, the network is trained by the idea of transfer learning, and the mixed data enhancement technique is used to expand the sample of data set. Compared with YoLoV4, the weight file memory is only 28.8%, the memory is only 48.1 MB, mAP is 10.65%, and the improved model is 97.06%.

Addressing the limitations of currently rare small target detection algorithms based on Human Visual Systems (HVS) that struggle with achieving satisfactory performance in complex backgrounds and lack high real-time capabilities, this paper introduces an innovative small target detection algorithm. This approach is founded on the integration of the Full-Direction Dilation Difference of Gaussians (FDD-DOG) and background estimation based on local contrast (LCBE). First, the high-frequency signal in the target area is enhanced by using the FDD-DOG operation, based on the difference between the real target and the background. Then, the background noise is further suppressed by assigning appropriate weights to the enhanced image by using local contrast-based background estimation. After that, the target to be detected is extracted by the adaptive threshold segmentation method. Experiments comparing the algorithm proposed in this paper with other HVS-based detection algorithms show that the method in this paper exhibits better detection performance in different infrared image sequences, where the background suppression factor is improved by an average of about 25 times and the signal-to-clutter ratio gain is improved by an average of about 15 times.

In this paper, an uncalibrated photometric stereo network model is proposed. Traditional photometric stereo has precise requirements on light source and object surface reflectance, which greatly limits the usability of photometric stereo, and the ideal network model should be able to handle arbitrary objects under uncalibrated light sources to complete the reconstruction. To solve this problem, we propose an uncalibrated photometric stereo neural network, First, an arbitrary number of images are input to the first-stage neural network system to estimate the light source information and complete the calibration to eliminate the dependence on the light source information. Then the calibration result is input to the second-stage neural network together with the image, and the multi-scale and attention mechanism is used to obtain the object surface texture information to achieve the surface normal reconstruction of the object. Our network (MAPS-Net) is compared with other uncalibrated photometric stereo methods on the DiLiGenT benchmark, and the average MAE achieves an excellent performance of 9.20.

In streak projection profilometry, how to reconstruct the 3D shape of an object quickly and accurately from a single streak pattern has been an important research area. Although the CNN-based fringe-to-depth method can directly reconstruct 3D from a single fringe, its accuracy is currently not as good as the traditional phase-shift technique. In order to improve the accuracy of 3D reconstruction, this paper proposes a U-Net-based global feature fusion network (GFU-Net), which introduces a global feature fusion module to fuse the global context information, and uses a feature fusion upsampling module to recover the spatial detail information, which solves the problem that it is difficult to accurately obtain the depth information of an object from a single fringe map. The experimental results show that the method proposed in this paper decreases the RMSE by 17.8% compared to the U-Net network, and the 3D reconstruction accuracy is higher and the error is smaller, which verifies the effectiveness and robustness of the method.

This study explores the application of large pre-trained language models in Classical Chinese translation, aiming to address the semantic understanding shortcomings of traditional methods. The LoRA-BLOOMZ model structure is employed, and experiments demonstrate that structured prompting along with the integration of LoRA technology and BLOOM model structure can enhance the performance of Classical Chinese translation. Through appropriate training strategies, stable training and optimized performance of the model are ensured, thereby achieving favorable BLEU scores in the Classical Chinese translation task. This study offers an effective new approach for Classical Chinese translation, providing a reference for the further application of large pre-trained language models.

In large scale landslide susceptibility assessment, it is necessary to reflect the differences of landslide formation background, conditions and evaluation indexes. However, the susceptibility assessment methods for small and medium scale area are not applicable. Sinan County has complex geological conditions, larger population and intense human engineering activity, in the event of a landslide, would be extremely costly. Thus, the site is a typical area for large scale and high precision susceptibility assessment. In order to explore locally adapted landslide susceptibility assessment method, 10 evaluation indicators, including altitude, slope, aspect, slope aspect, lithology, distance to fault, TWI, NDVI, distance to road and land types, are selected to evaluate landslide susceptibility, through Logistic Regression model (LR), Decision Tree model (DT) and Random Forest model (RF). Experimental results of model evaluation using receiver operating characteristics (ROC), area under the curve (AUC) and accuracy (ACC) showed that the RF (ROC=0.959, ACC=0.897) were more accurate than DT (ROC=0.925, ACC=0.884) and LR (ROC=0.786, ACC=0.703). High and very high landslide-prone areas are mainly concentrated in the steep slope of the valley bank, the geological environment is fragile, and the human engineering activities are strong. The results from this study demonstrates that the RF can identify landslides well in large-scale landslide susceptibility assessment, and provides scientific basis for disaster prevention and mitigation in large-scale landslide susceptibility assessment of mountain towns.

At present, there are some problems in traditional image dehazing algorithms, such as prior knowledge limitation, color distortion, high computational complexity and inability to process multi-scale information. This paper proposes an image dehazing network based on multi-scale lifting and attention mechanism fusion U-Net structure. Firstly, the downsampling operation is used to obtain multi-scale feature maps, and the feature maps of the encoder and decoder are connected by jump connection to realize feature fusion. The purpose is to transfer and fuse information between feature maps of different scales to improve the performance of the model. At the same time, in order to better aggregate cross-window information, an overlapping cross-attention module is introduced to enhance the information complementarity of adjacent window functions. By comparing the proposed algorithm with the other five algorithms in the public datasets Reside and NTIRE, the experimental results show that the proposed algorithm performs well in defogging effect and color retention. Specifically, the average peak signal to noise ratio (PSNR) of the proposed algorithm reaches 27.5 dB, and the average structure similarity (SSIM) is 0.9573. Compared with the other five algorithms, the performance has been significantly improved. Compared with the best comparison algorithm, the PSNR of the proposed algorithm is improved by 2.17 dB, and the SSIM is improved by 0.0063. These experimental results fully prove that the proposed algorithm has excellent performance in dehazing performance and color retention.

During the process of actual tasks, automated guided vehicle will meet some problems, such as incomplete real-time monitoring, off-track, etc. An AGV simulation system based on interactive control is designed. Firstly, the visual system guides the real AGV for operation. Meanwhile, the simulation system establishes the virtual AGV motion model according to the driving rules of the real AGV. By obtaining and processing the driving data of the real AGV, the operating situations of the real AGV are simulated in real time in simulation system. Finally, the algorithm of track correction detects the operation status of the real AGV and performs remote synchronous correction when the real AGV deviates from the task track. The experimental results show that the simulation system and the algorithm of track correction are real-time and effectiveness, which provides technical support for other similar AGV simulation systems.

Data acquisition and processing of existing buildings are the foundation for virtual reconstruction. However, due to the complexity and diversity of existing buildings, data acquisition and processing often face certain difficulties. In order to improve the accuracy and efficiency of virtual reconstruction of existing building restoration, the application of 3D laser scanning technology in virtual reconstruction of existing building restoration was studied. Using 3D laser scanning technology to extract point cloud data from existing building images, and using Laplace filtering algorithm to filter and process point cloud data from existing building images. On this basis, virtual reconstruction of existing building restoration is achieved by integrating the pixel translation vector of existing building images, the optimal rotation matrix of point cloud data, and the similarity transformation scaling factor. The experimental results show that the proposed method can effectively achieve virtual reconstruction of existing building restoration, has good virtual reconstruction effect of existing building restoration, and can effectively improve the efficiency of virtual reconstruction of existing building restoration.

Recent studies have shown that players' pregame emotions can be an important data source when predicting NBA players' game performance, and that sentiment states have an impact on athlete's game performance. Our project aimed to determine whether a similar phenomenon could be observed in professional tennis players: sentiments states influence game performance. To do this, we use web crawlers to obtain tennis match data, track the tweets of participating tennis players, collect tweet content from posts posted on their Twitter accounts. We used natural language processing for sentiment. We used the VADER sentiment analysis tool to sentiment-analyze the content of their tweets. In this way, they can identify their emotional state before the game and count the number of correlations to determine whether their emotional state is related to the performance of the game. Through experiments, we can conclude that the emotional state of athletes is closely related to their sports performance. Next, we combine the sentiment analysis data with other behavioral data of the athletes to build a model that predicts the outcome of the game based on a decision tree algorithm. Using the model to predict the performance of players in the game with an accuracy rate of 85%, it provides guidance for tennis players' teams in future training plans and game plans.

With the development of machine vision, deep learning-based target detection methods are crucial in maritime ship detection. Although some recognition methods, such as YOLO (you only look once) and SSD (single shot multibox detector) have achieved good results, there are still problems with low detection accuracy and easy missed detection and false detection for small moving targets at sea. This research proposes an optimized YOLOv8 algorithm with an attention mechanism to enhance feature extraction and fusion, and adds a detection head for small targets at sea. Then data augmentation methods are introduced. Finally, videos are recorded on the laboratory's visual simulation system and a data set is constructed. Experimental findings reveal that the optimized method 's mAP@0.5 on the self-made data set is 2.9%, higher than the original method, and the mAP@0.5 of the speedboat small target is improved by 9.5%.

It is crucial to monitor and predict the trajectory of ships for preventing maritime accidents and enhancing maritime efficiency. Traditional data mining algorithms face serious challenges in terms of computational complexity and generalization capability. To address these challenges, a deep learning-based ship trajectory prediction method is proposed in this paper. By performing clustering on trajectory data, similar trajectories are grouped into clusters, effectively reducing the complexity of the data. Subsequently, the Seq2Seq model is employed to model and predict trajectories within each cluster. Experimental results demonstrates effectiveness and superiority of the proposed method in ship motion trajectory prediction tasks.

During the process of scanning the shape of a large object, multiple scans from different angles are required due to the influence of surface complexity and scanning field of view, and then these data are stitched together into a whole. This paper proposes a tracking-based optical measurement point cloud stitching method based on circular coded points. During the scanning of the object by the structured light scanner, the binocular tracking camera real-time locates and pastes the circular coded points on the scanner. Using this marker as an intermediary, the reconstructed 3D point cloud in the coordinate system of the structured light scanner is transformed to the reference coordinate system of the positioning and tracking binocular camera to achieve point cloud stitching and complete the reconstruction of the three-dimensional morphology information of the object being measured.

With the widespread development of machine learning and deep learning technology in recent years, artificial intelligence technology has been widely applied in various fields. Traffic flow prediction has been a hot research direction in recent years. Traffic flow is an important influencing factor in the field of transportation and travel, playing an important role in the rational allocation of transportation resources and ensuring the smooth performance of transportation and travel. It is necessary to predict future traffic flow based on historical travel data. Some existing deep learning methods require high time complexity and high hardware costs, and that have some shortcomings in prediction ability. This article proposes a CATformer model combining convolutional neural networks and attention mechanisms to solve the traffic flow prediction problem in time series problems, extracting and fusing features from multiple vector spaces for traffic data, The experimental results of predicting the flow of future traffic nodes show that the CATformer model has improved prediction accuracy compared to the benchmark method, achieving the task of traffic node flow prediction based on time series.

The short-term HRV signal before the onset of PAF for 48 subjects is analyzed, the HRV signal preprocessed is first separated by wavelet transform, and then the bispectrum analysis and feature extraction are performed on the harmonic and 1/f fractal signal of separated HRV. Compared with the control group, the bispectrum amplitude of HRV harmonics and 1/f fractal signal increase significantly in the ten minutes before the onset of PAF, and more high-frequency components appeared in the harmonic signal, its Mave, H1, H2, and H3 features change significantly (p<0.05); while the features of HRV harmonics and 1/f fractal signals 5 minutes before the onset of PAF change more significantly (p<0.0005). This indicates that the imbalance of the autonomic nervous system function is aggravated when approaching atrial fibrillation, and suggests that certain bispectrum features may be sensitive indicators for predicting the occurrence of PAF.

To ensure the safe operation of subcritical extraction plants, the temperature of the subcritical extraction vessel must be predicted. Therefore, a time series prediction method for the temperature inside the vessel based on the Informer model is proposed. First, time series data generated during the operation of the subcritical extraction vessel is collected. Then, the Informer prediction model is built using the time series data to predict the temperature of the subcritical extraction vessel. Finally, the curve of the data predicted by the model is compared with the actual data and the prediction results are evaluated to demonstrate the feasibility of this method. The comparison of the prediction results with the LSTM and GRU models shows that the Informer model has better prediction performance.

In order to solve the problem that it is difficult to quickly detect the pass rate of the reed high vertical sand barrier after it is made, a method for the pass rate detection of the high vertical sand barrier of reed based on K-means clustering was proposed. The images of the reed high vertical sand barrier were collected and preprocessed, the K-means clustering algorithm was used to cluster the reed bundle area in the image, and the separation of the background and the reed bundle area in the image was completed by the binarization operation. Canny edge detection and minimum circumscribed rectangle operation were performed on the binary image, the midline of the smallest circumscribed rectangle was obtained, and the distance between the reed bundles was calculated with the help of Euclidean distance. Experimental results show that this method can effectively detect the gap of the reed high vertical sand barrier, improve the speed of pass rate detection, and provide a new technical means for desert control.

In order to monitor the traffic flow on the road surface, an experimental process of automatic road network extraction and vehicle monitoring using high-resolution remote sensing images was designed. Firstly, the D-Link network is used to perform binary semantic segmentation on the high-resolution remote sensing images to extract the road area, the coherence, edge expansion and smoothness of the extracted road are optimized by morphological closed operation processing for the preliminary extraction results, the road area is obtained by using the extracted road area as a mask, the vehicle target detection is carried out by using the deep learning algorithm, and the vehicle target density is represented by drawing the heat map. The experimental results show that the designed experiment can realize the monitoring of road traffic flow.

Due to the complex scenes of reality in the scrapyard, the overlapping between scrap metals causes occlusion, resulting in missed detection and false detection problems during the classification and grading of scrap metals. In response to this problem, we propose a network model for scrap metal classification and grading, named FG-Net, which improves Yolov5s. First, we replace the convolutional layers in the backbone with the proposed MBCC module, so that the backbone network can adopt strategies such as depth-wise separable convolution, residual connection, channel and spatial expansion, which not only improves the feature extraction ability, but also enhances the efficiency and accuracy of the network. In the feature fusion stage of this model, we introduce the RFB module to improve the receptive field and increase the flexibility of feature fusion, enabling the model to fuse more feature information. Finally, an improved loss function is used to address the missed detection and false detection issues caused by the closeness of scrap metals.This paper first verifies the model on the official datasets PASCAL VOC(2007+2012) and COCO, with mAPs of 84.6% and 59.8%, respectively, which are 2.1% and 3% higher than Yolov5s. The accuracy of the self-made scrap metal dataset reaches 87.4%, with an mAP of 86.1%, which is 10.1% and 10.4% higher than Yolov5s, respectively. Through experiments, it is proved that this model not only improves the accuracy of scrap metal recognition, but also solves the occlusion problem in the current scrap metal classification and grading process.

At present, the rapid development of Internet information technology has brought convenience to education and teaching. E-learning based on network has become an important way of learning. However, the traditional E-learning learning environment usually only provides course learning resources by the fixed preset program, which is not friendly to the user's personalized learning experience. Intelligent learning system (ILS) is a support environment that can dynamically provide personalized and intelligent learning services by diagnosing the individual differences of learners' users. It mainly provides current users with suitable learning paths or learning resources via personalized recommendation technology, so as to meet users' learning needs and improve their learning experience. However, the current research on intelligent learning system is still in the exploratory stage. Many researchers have studied intelligent learning system from different perspectives, but their research achievements still need to be explored and improved on the effect of intelligent recommendation. Based on this, this paper will further explore the key technology of personalized recommendation of intelligent learning system on the basis of the related research case analysis. In order to improve the recommendation accuracy of intelligent learning system, we will model and analyze the user behavior characteristics, and further explore the recommendation solution of personalized learning path of the system. The simulation experiment shows that the research results have a good personalized recommendation effect, which can effectively diagnose and analyze the current user's learning state characteristics, and then dynamically provide a suitable personalized learning path to meet the user's learning needs.

As the global population grows and food demand rises, the efficiency and automation of agricultural production have become one of the important challenges facing the world today. Among many crops, tomatoes occupy an important position in modern agricultural production due to their widespread cultivation and consumption. However, the traditional tomato growing and picking process suffers from inefficiency, high costs, and over-reliance on manpower. In response to these challenges, this article proposes a solution using advanced deep learning technology, which is to apply the YOLOv8_GAM_BiFPN network for tomato target detection.

In the task of multi-person tracking, we have observed that ID switches in applications can lead to confusion in tracking targets, significantly reducing the efficiency of target tracking in actual work and the user experience. Therefore, this paper proposes a new method for multi-person tracking, building upon the BoT-SORT, with a focus on identity switching. The proposed framework integrates spatial location constraint and occlusion awareness into the tracking process. By combining spatial location and occlusion for track initialization, and methods for ID switch verification and rectification, we have enhanced the utilization of low-confidence detections, effectively reduced ID switches in multi-person tracking. The effectiveness of the proposed method has been verified through ablation experiments and validated on the MOT20 benchmark, The HOTA increased by 0.3 to 63.6, the MOTA increased by 0.3 to 78.1, and the IDF1 increased by 0.9 to 78.4. Particularly, the number of ID switches decreased from 1313 before the improvement to 1007, a reduction of 23.3%.

Laser interference can interfere with target detection systems to achieve the goal of protecting the target. For example, it can obscure the target and change information such as brightness and contrast in the surrounding region of the interference zone. Therefore, it is necessary to analyze the impact of laser interference on target detection algorithms and evaluate the anti-interference performance of the algorithm. This paper analyzes the impact of laser interference on two target detection algorithms, Faster-RCNN and YOLO-V3, from two perspectives: target occlusion rate and target similarity. Then, a target-oriented method for dividing the effective region of laser interference (TODERLI) is proposed. The effectiveness of the algorithm is verified through experiments.

Light distribution within fruit tree canopy is one of the important factors to optimize fruit tree architecture and to improve the potential production and fruit quality. Take the canopy structure, radiation environment and light transmission into consideration, plant model simulation analysis is an important way to measure light distribution within tree canopy in the field. There were 700 to 900 multi-view visible images for one Chinese bayberry tree obtained by moving a camera around the tree canopy. Using machine vision method, three-dimensional point cloud was reconstructed from images based on the multiple-view stereo and Structure From Motion (MVS-SFM) algorithm. The encapsulating, smoothing and filling algorithm were used to transform point clouds into smooth surfaces consisting of triangular facets in computers automatically. Canopy height, canopy width, branch length and circumference for the first three grades were measured in the orchard to evaluate the performance of the 3D structural model. Ray-tracing method in a 3D radiation model was used to calculate light distribution within the 3D tree canopy. The light distribution result within the 3D structural model was visualized using computer technology. The results showed that the multi-view image method well reconstructed the tree canopy. There was satisfactory correspondence between acquired and measured tree trunk height, circumference and canopy width, with the root mean squared error (RMSE) smaller than 1.2 cm and the consistency coefficient (d) larger than 0.90. The light was mainly distributed in the periphery of the canopy and in the upper-middle layer of tree canopy. Due to the sever occlusion of foliage and branches, the light was less in the inner and lower layer of the canopy. The fraction of light interception was 0.10, 0.17 and 0.05 for the bottom, middle and upper 1/3 canopy layer respectively. This study provide possibility for high-throughput 3D canopy structure measurement in the orchard and provide a reference for tree pruning, ideal tree structure design, and fruit yield and quality improvement.

Aimed at the problem of bearing surface defect detection, a method based on context-aware saliency detection is proposed. Firstly, the image of bearing surface defects was preprocessed, and homomorphic filtering method was used to weaken the influence of uneven illumination and environment. Then, the saliency of the defect area was enhanced by the context-aware visual saliency detection method, and a saliency map highlighting the defect area was generated. Finally, Otsu method was used to segment the defect of saliency map, and the defect detection results were obtained. The results showed that the context-aware saliency detection method can accurately detect all kinds of defects on the bearing surface, and has better precision and recall than other significant detection methods.

Current black-box adversarial attacks have proven to be highly effective in generating adversarial texts that can successfully deceive natural language processing models, thereby revealing potential weaknesses in these models. This research proposes an innovative transfer-based black-box attack method, which capitalizes on the combined generative and discriminative abilities of the diffusion model. To ensure semantic similarity and enhance the adversarial ability of generated texts, well-designed semantic-preserving and adversarial objectives are introduced to the training procedure of the diffusion model. The results show that the proposed method can generate adversarial texts that successfully attack text classification models.

Aim to solve the problems of high-cost and low-efficiency in manual blueberry picking under harsh natural environment, an automatic blueberry identification using machine-learning YOLOv5 (You Only Look Once) algorithm and CIE saturation enhancement is proposed. The color contrast of blueberries between different ripeness is enhanced by saturation and desaturation processing in the CIE Lu'v' color space. The augmented image dataset is then trained by YOLOv5 to obtain a target detection model for both ripe and unripe blueberries. Experiments are conducted by comparing the original YOLOv5 framework detection and the augmented YOLO detection results. By CIE saturation enhancement, the identification precision in training, recall and the mean average precision are found to be increased with 3.0%, 2.0% and 2.6% respectively.

During the reconstruction process, the measured object is affected by its own reflection, environmental illumination, and other factors, resulting in a large number of mismatching points during the stereo-matching process, thus affecting the accuracy of the three-dimensional reconstruction. Aiming at the above problems, this paper adopts a scanning method combining single-line laser and binocular vision. The binocular epipolar constraint is used to obtain the initial matching points, and the optical plane constraint is added to eliminate the false matching points. The experimental results show that the 3D modeling error of the workpiece is maintained within 0.5cm, which verifies that the method has certain feasibility and good accuracy in solving problems such as mismatching.

For the YOLOv8 pedestrian detection issue on embedded devices with high computational complexity and deployment challenges, we propose a novel lightweight pedestrian detection solution. This involves replacing YOLO's backbone network with lightweight models, MobileNetV3 and EfficientNetv2, and modifying the attention mechanism in the model by introducing the CBAM attention mechanism. Additionally, an SPPF module is added to the last layer of the model to enhance feature extraction. To achieve lightweighting in the head segment, we halved its parameter count. Experimental results show that compared to directly replacing the network, our approach successfully reduces the parameter count while achieving higher accuracy.

With the continuous progress of society, people's lifestyle and quality of life are constantly changing, although the overall life expectancy of human beings has been extended, but lung diseases are still a serious threat to human health. In order to improve the object detection performance of lung nodules, the Fast R-CNN model was improved. By adding the CBAM attention mechanism module to the convolutional layer of the ResNet-50 and VGG16 feature extraction models, the model focuses more on the key feature extraction in the image. Due to the small size of pulmonary nodules in CT images, a object size selection box based on statistical data was introduced, and the improved Fast R-CNN-VGG16- CBAM model was used in the experiment with VGG16 as the backbone. Compared with the original model, the precision is increased by 6.01 percentage points, the recall rate is increased by 3.34 percentage points, and the mAP value is increased by 4.54 percentage points, and the improved Fast R-CNN-Resnet50-CBAM model is improved by 2.71 percentage points, the recall rate is increased by 4.6 percentage points, and the mAP value is increased by 4.03 percentage points. In addition, compared with the first-stage detection model YOLOV3, the improved Fast R-CNN-Resnet50- CBAM model has 5.2 percentage points higher precision and 13.36 percentage points higher mAP value. Experimental results show that the designed Fast R-CNN-Resnet50-CBAM model can effectively improve the performance of lung nodule object detection.

Time series usually have certain trends and inherent patterns, so their future values can be predicted based on current information. The time series forecasting with high accuracy can provide decision basis for the relevant personnel, and then improve the efficiency of decision making. There are more and more researches on models for time series forecasting. In view of this, this paper proposes a hybrid MODWT-PSO-SVM model for time series forecasting. It is found that the constructed hybrid model can achieve better predictive performance. Furthermore, this paper investigates the impact of decomposition levels and the form of wavelet functions on the predictive performance of the constructed hybrid model. Research has shown that the predictive performance of the constructed hybrid model tends to decrease first and then increase with the increase of decomposition levels. This means that the predictive performance of the hybrid model varies with the number of decomposition layers, and an optimal number of decomposition layers can be found to achieve the best predictive performance. In addition, different forms of wavelet function also affect the prediction effect of the mixed model. When the Sym4 wavelet is used, the prediction effect of the hybrid model is the best. While when using Coif3 wavelets, the hybrid model has the worst prediction performance. Therefore, the predictive performance of the hybrid model can be improved by adjusting the number of decomposition layers and changing the form of wavelet functions. This research can provide a model reference for time series forecasting and supplement the existing research.

This paper proposes an improved GAN-based super-resolution (SR) network to address the issues of detail loss and low feature utilization in remote sensing image SR reconstruction. First, we propose a residual block (ERDB) containing multi-scale receptive fields (MRFs) to fully capture features at different scales, and use the attention module to dynamically adjust the features. In the discriminator network, we use the relative discriminator to compute the relative probability instead of the absolute probability and incorporate an attention module to help generate images containing more texture details. Experimental results demonstrate that our method has improved objective evaluation metrics on both UC Merced and NWPU-RESISC4 datasets and achieved good visual effects.

Aiming at the problem that the discriminator of original conditional generative adversarial network cannot classify real samples, a semi-supervised generative adversarial network structure is proposed to realize the classification function of real samples. Considering that the traditional semi-supervised generation adversarial network is too rough to evaluate the generator, the loss function during the training of the generator is improved, and the generator can generate more realistic samples. The mature discriminator corresponding to the improved semi-supervised generation countermeasures network is deployed in the vibration monitoring terminal, and the high accuracy is obtained, which can effectively identify the hidden danger information of slope when the static, vehicle passing and impact events occur.

Cow face identification plays a crucial role in the cattle management system. Previous studies have primarily focused on radio frequency identification, and a few researchers devote to the cow face identification field. In this paper, instead of solely extracting features from individual images, we have constructed datasets for cow face identification. The datasets include the facial images of an all-black cow, an all-white cow, and a mixed black-and-white cow. We apply the convolutional neural networks method by utilizing ResNet backbone architectures, and additionally, we incorporate different loss functions and attention modules to enhance the model’s capacity. The results demonstrate that our methods have achieved an identification accuracy rate of 97.04% and FRR of 5.06%, which also improves identification speed and performance compared to other studies, marking a notable advancement in cow face identification.

In response to the insufficient feature information of smaller electronic components in the current YOLOv3 algorithm used in object detection, which leads to false detections, missed detections, and low average accuracy, this paper proposes an improved attention mechanism-based YOLOv3 algorithm for small object detection. Firstly, the model introduces a category weighting factor α to overcome the problem of imbalanced training samples, enhancing training robustness and improving classification accuracy. Secondly, an attention mechanism is introduced to learn the importance of each channel's characteristics, enabling weight allocation for primary and secondary information, suppressing secondary information, and eliminating some redundant information, effectively improving the learning direction of the model. Lastly, by incorporating the category weighting factor α and L2 regularization term into the loss function, the training of the network model can converge faster and more stably. The proposed algorithm achieves an average precision of 92.86% for resistor recognition, a 1.07% improvement compared to the classical YOLOv3 algorithm. For chip recognition, the average precision is 97.23%, an improvement of 0.8% compared to the classical YOLOv3 algorithm. These results demonstrate a certain enhancement in the detection performance of small objects.

To realize the automatic recognition of oracle bones, this paper presents an oracle bone recognition technology based on the development of Chinese characters. In the experimental process, the model of YOLOv7 algorithm is used to realize the recognition function of oracle script. The model of YOLOv7 algorithm is improved by adding the Coordconv structure and replacing the traditional NMS (Non-maximum Suppression) with the matrix NMS, which effectively improves the recognition accuracy and speed of the oracle script. Moreover, in this work, oracle recognition is introduced based on the semi-supervised learning of training, learning, Jinwen, the relationship between the seal script and the clerical script, and more script features are used to search existing words. In the experiment, the recognition accuracy of the oracle script reached 81.12%. This also proved that there is a certain relationship between the oracle script characters in the process of evolution, which is a basis for the research of oracle script characters based on the evolution of Chinese characters.

In order to determine whether electric vehicle drivers are wearing helmets, we propose a study of an improved YOLOv5 electric vehicle helmet recognition algorithm to detect whether an electric vehicle driver is wearing a helmet or not. First, by using the triple attention mechanism to capture interactions across dimensions in the backbone network, the target feature region can be localized more precisely. Then, by using the Focal-EIoU loss function to enhance the model training, the model's performance can be improved more effectively. Through experiments, it is shown that the enhanced YOLOv5 algorithm achieves a mean average precision (mAP) of 92.48%, which is 9.88% higher than the basic YOLOv5 algorithm. This accuracy has risen to 95.24%, meeting the requirements for recognizing the helmet-wearing condition of electric vehicle drivers on the road.

The technology of adaptive deformation compensation for thin-walled parts, based on on-machine measurement (OMM), has found widespread application in various industries. However, due to the size of the ruby probe, a certain amount of space must be reserved along the workpiece boundary when using the probe to sample the surface of thin-walled parts. This precaution is taken to prevent any collisions or overtravel incidents involving the probe. Consequently, this reservation results in an inability to gather information from the boundary area of thin-walled parts. After undergoing multi-layer milling compensation processing, stepped processing transition areas can develop. In light of the theory of NURBS (Non-Uniform Rational B-Spline) surface reconstruction, this article introduces an algorithm that extends a distance along the tangent direction at the beginning and end of the sampling curve to expand the boundary. A sampling surface was constructed and expanded, and the maximum error γ between the two surfaces was calculated to be 0.27×10^-3mm. The results showed that it met the accuracy requirements. The verification results confirm that this method can indeed be practically applied in the field of adaptive deformation compensation technology for thin-walled parts.

Research on luggage of passengers has increased recently in the field of security management in transportation hubs. Previous research primarily focused on the impact of luggage on crowd evacuation in the case of emergencies. How to recognize a luggage, and how they impact the movement of crowd still worth further and systematic study. The abnormal states of luggage (e.g., falling off, tail flicking and acceleration) can perturb the natural march of crowd. This study proposes an approach of luggage falling-off recognition involving computer vision technology. This approach considers the movement of passenger and luggage as a combined motion. Experiment results indicate that the proposed approach in this study can recognize the falling-off state of luggage effectively and can provide a practicable technique reference for carried object analysis in transportation hubs.

Due to the close similarity in tire textures, it is challenging to predict the type of a tire with a single deep learning model. Therefore, an ensemble learning based tire classification method is proposed in this paper, which leverages the advantages local perception in Convolutional Neural Networks (CNNs) and global attention in Vision Transformers (ViTs). Initially, individual networks based on CNNs and ViTs are independently trained. Subsequently, the backbone networks of these models are frozen, and the features obtained from different models are concatenated. The concatenated features are then fed into a feed forward network for prediction. The experimental results demonstrate that the proposed ensemble model exhibits higher accuracy and stronger generalization capability in tire classification tasks. Compared to individual predictions with single model based on CNNs and ViTs, the accuracy is improved by more than 18.51% and 8.05%, respectively.

With the booming development of the internet industry and the entry of big data into people's lives, it is particularly important to quickly extract people's needs from massive text information. This article proposes a text classification method based on an improved long and short term memory network. To address the problem of being unable to encode information from back to front, the Bi-LSTM model is improved to concatenate forward and backward outputs with the earliest word vectors to obtain the final word representation, and to find the most important word for classification through the maximum pooling layer, thereby improving the performance of text classification. The experimental results show that the algorithm proposed in this paper has better classification accuracy than other classification algorithms when the number of iterations is large.

Currently, AI for data image processing has been widely studied in various fields of geology; such research has been greatly improved in terms of timeliness and accuracy, but no AI application with good results has been proposed for the recognition of pores and pyrite in shale images imaged by electron microscopy. To address the problem of intelligent recognition of pores and pyrite in shale images imaged by scanning electron microscopy, this paper proposes to apply Resunet, a novel convolutional neural network combining Resnet and Unet, to the intelligent recognition of pores and pyrite in shale images, and at the same time train the Resunet model through the dataset, evaluate and test the model, and obtain the resU-net semantic segmentation model The correct rate of recognizing pyrite and shale pores reaches 96.42%; the experimental results show that resU-net has the highest recognition accuracy compared to other models in the task of recognizing pyrite and shale pores on scanning electron microscope images, and this study effectively improves the problem of intelligent recognition of pores and pyrite in shale images imaged by electron microscope.

Rice pests can adversely impact the quality and productivity of rice, and proper identification of pests is beneficial for growers to taking prompt steps to manage pests. Although the current crop pest identification model has achieved a high recognition accuracy in recent years with the application of deep learning, it is challenging to deploy on the edge devices with limited computing resources and low performance, and it cannot meet the needs of practical agricultural production. Therefore, we propose an effective real-time lightweight pest detection method, Mobile-CANet, which is based on the convolutional neural networks. This study develops a new network structure to enhance the feature extraction capabilities in order to address the issue of location information being lost during feature extraction. The capacity to extract features is enhanced by this network structure, which also incorporates transfer learning to enhance generalization ability. The results of the experiments show that this method's accuracy and number of parameters are superior to those of the existing picture classification models. On the publicly available dataset of rice pests, it can reach 85.3% accuracy with only 3.312M parameters, which can offer technical assistance for the implementation of mobile algorithms in the future.

Chinese medicinal materials play an important role in the health care of our people, and the quality of Chinese medicine has received more and more attention. The quality of Chinese medicinal materials and the authenticity of drugs are related to the therapeutic effect, so the identification of authenticity of Chinese medicinal materials is very important. China is very rich in traditional Chinese medicine resources, and many kinds of traditional Chinese medicines have brought great difficulties to the identification of traditional Chinese medicinal materials. The characteristics of traditional Chinese medicines are not only the appearance characteristics of medicinal materials, but also the external manifestations of the internal organizational structure of medicinal materials and the chemical components contained in them. In research, the essence of the trait identification of traditional Chinese medicine is to explain the correlation between trait characteristics and internal tissue structure and its intrinsic chemical composition in modern science. At present, some new methods and technologies are applied to the identification of traditional Chinese medicine traits, such as simulating the electronic nose and electronic tongue generated by the human olfactory organ and taste system, and image recognition and identification are also used. These new methods and technologies are all extensions of trait identification and are conducive to the development of Chinese medicine.

Interactive web interfaces are widely used, yet users often struggle to discover important and useful information. Psychologists have pointed out that in the process of human visual perception, the global topological characteristics of objects will be perceived first. In this study, we propose a cognitive experiment system to explore what object properties affect visual efficiency in interactive web environments. Our experimental results showed that, first, if a small number of colors are allocated to different objects, these objects could be more effectively memorized than no color or too many colors. Second, in static single-page, different topology structures will not cause apparently different visual attraction, however, topological structural differences could be stronger memorized. Third, in dynamic HTML, the motion changes will induce stronger visual attention and memory intensity, than their own topology changes. Subsequently, we verified the feasibility of the experiment system in real webpages and demonstrated that the above results can theoretically guide better web interaction design works with less cognitive load of users, higher operation efficiency and more natural user experience. In addition, some design reference suggestions are also provided in our study.

Named Entity Recognition (NER) in classical Chinese is a crucial task in the natural language processing of classical texts and the construction of knowledge graphs. Its primary focus is the extraction of entities such as events, locations, and personalities from historical documents. Current mainstream NER methods predominantly utilize single-level textual features, neglecting the character structure and character-word features, which limits the acquisition of sufficient character structure and word information. Consequently, this study proposes a novel NER method for classical Chinese, integrating character and positional multi-feature vectors. Initially, a multi-feature word embedding method combining character and positional information is established. Text sequence features are then extracted using a Bidirectional Long Short-Term Memory (Bi-LSTM) network. Subsequently, a multi-head attention mechanism is employed to capture context-sensitive features across various subspaces. Finally, a Conditional Random Field (CRF) outputs the predicted annotation sequence. Experimental results on the C-CLUE dataset demonstrate that the proposed method significantly enhances NER performance compared to baseline algorithms like Bert, achieving high average F1-Scores of 77.04%, thus exhibiting high accuracy and usability.

Large language models (LLMs) are increasingly vital across diverse applications. This study delves into the utilization of the Alpaca language model, a cost-effective and high-performing variant, for sentiment analysis. The research systematically evaluates the Alpaca model's performance across a range of datasets, with an emphasis on its ability to excel in sentiment analysis tasks across varying domains. The study sheds light on the practical utility of the Alpaca model in sentiment analysis applications, providing insights into its adaptability and effectiveness. Furthermore, this investigation extends its inquiry into the realm of consumer-grade graphics, examining how large language models perform in sentiment analysis within this context. Notably, the study uncovers that these models demonstrate proficiency in nuanced tasks with limited answer options. Moreover, the research underscores the potential for improved accuracy through the utilization of larger training datasets. In conclusion, this research showcases the Alpaca model's efficacy in sentiment analysis and advances our understanding of LLMs capabilities in this field. It underscores their potential to enhance context-aware and precise sentiment analysis solutions, offering valuable contributions to the broader landscape of natural language processing.

As vein identification technology advances, more and more vein recognition products are coming into the public's view. Additionally, vein recognition technology boasts a much higher level of security than previous recognition technologies. The convolutional neural network model is frequently used in vein recognition research to increase vein recognition accuracy. However, because the model learns the noise in the image, vein recognition accuracy cannot be further increased. To tackle the aforementioned issues, a neural network model based on Canny is suggested for the detection of palm veins. The input image's feature information was extracted using the convolutional layer, controlled by the Canny layer to learn the vein image's texture features. The self-attention layer was then utilized to increase the convolutional layer's robustness when learning the vein image's feature information. Based on the experimental results, it can be concluded that the Canny-based vein recognition neural network model performs better at extracting texture features from vein images and avoids the network from learning unnecessary feature information. It has also demonstrated improved accuracy and equal error rates on three publicly available palm vein datasets: the CASIA-PV200 palm vein dataset (94.33% and 5.83%), the TJU-PV600 palm vein dataset (94.88% and 4.86%), and the VERA-PV220 palm vein dataset (95.64% and 4.86%).

In the era of big data, power companies have accumulated a large number of policy documents at the government, industry, and company levels. Structuring these policy documents, efficiently extracting their core content, and fully exploiting their value is of great significance for optimizing business models, supporting management decisions, and building digital services in power companies. With the development of artificial intelligence technology, knowledge graphs have emerged. Knowledge graphs can link fragmented knowledge through relationships, building structured document knowledge bases, making the knowledge contained in the documents easier to understand, query, manage, and apply.

Recent security incidents against the global DNS ecosystem show a close association with outdated versions of DNS servers. Attackers and security researchers widely employ server fingerprinting as a pre-exploit technique. Considering that a single active or passive fingerprinting technique is difficult to achieve accurate and effective results, we proposed a fingerprinting method combining active and passive techniques to identify different DNS servers. We record DNS server fingerprints by collecting interaction borderline behavior of different DNS servers. After that, we propose a novel hierarchical fingerprinting model named as fpdns-ng algorithm. We apply machine learning models like Decision Tree, KNN and Multilayer Perceptron to classify different kinds of DNS servers. Experiment results prove the effectiveness of our method and show that our approach is more accurate than the state-of-the-art in fingerprinting mainstream DNS servers.

In recent years, human pose estimation has seen substantial advancements. However, occlusion remains a persistent challenge, leading to issues like missing keypoints, ambiguity, and abnormal poses. This paper introduces a novel framework, comprising two core components: the Initial Pose Net and the Multi-Scale Structure-Aware Semantic GCN (MSS-GCN). Our approach starts with the Initial Pose Net, which predicts an initial pose with the aid of attention mechanisms, enhancing feature precision. Subsequently, the MSS-GCN refines this initial pose to yield the final pose. The MSS-GCN is a network consisting of multiple parallel multi-scale subgraphs, informed by the human body’s geometric constraints. It leverages prior knowledge of human anatomy to learn constraints regarding keypoints and uses feature information to compensate for the absence of semantic data in coordinates. The MSS-GCN excels in capturing the structural aspects of the human body and effectively handles the challenge of long-distance keypoint predictions. Additionally, our module is designed for easy integration into other networks, enhancing its versatility. Through extensive experiments on standard benchmark datasets for occluded human pose estimation, we demonstrate that our method surpasses existing state-of-the-art approaches. This work marks a significant stride toward enhancing pose estimation accuracy in occlusion scenarios.

The automotive advanced driver assistance systems (ADAS) are increasingly sophisticated due to the rapid advancement of intelligent connected vehicle technology. Consequently, higher demands are placed on both hardware infrastructure and onboard systems. To tackle the challenge of achieving complex traffic object recognition on limited hardware resources, a novel multi-sensor feature-level fusion strategy is proposed, which integrates data from millimeter-wave radar and visual sensors. By achieving synchronization of different sensors in both time and space domains, this strategy enhances the accuracy and robustness of traffic object recognition, while also improving computational efficiency and overall system performance.

In response to the issues of low recognition accuracy in silent electromyographic facial action speech reconstruction tasks, this paper combines residual network (ResNet) and Transformer model to design a Res Transformer model based on ResNet structure and Transformer network. The model consists of three connected ResNet structures and several Transformer modules, and features are extracted using ResNet, The Transformer structure converts the electromyographic signal into a Mel frequency spectrum of 80 frequency bands, and sends the Mel frequency spectrum into the HiFiGAN network for speech reconstruction, ultimately obtaining the audible speech signal under silent action. In addition, our work also integrates acoustic speech signals, extracting the Mel frequency spectrum of synchronous acoustic speech signals as features, fully utilizing the time-frequency domain features of electromyography and speech signals. The experimental results show that the proposed Res Conform model has a phoneme recognition rate of 91.86% and a word error rate of 25.6%. Compared with using structures such as Transformer and LSTM, the Res-Conformer model has achieved effective improvement in recognition accuracy.

Point cloud registration involves aligning two point clouds with different spatial perspectives, commonly used in computer vision and artificial intelligence. Traditional methods lack robustness and accuracy, while deep learning approaches require many parameters, making them computationally expensive and time-consuming. This study aims to develop a compact and efficient model using knowledge distillation techniques to address challenges like initial pose differences and incomplete overlap, improving registration accuracy. Experiments on the ModelNet40 dataset with noisy and partially overlapping point clouds show that the distilled small model achieves favorable registration outcomes with fewer parameters and training time, effectively addressing the overlap problem.

With the rapid development of Internet social platforms, some extreme remarks, online rumors and negative thoughts often appear in various news, Weibo, WeChat, and promotions, which seriously affect social stability, stability and unity. How to dig out public opinion information from these media information texts and effectively monitor and manage online public opinion is of great significance. Text sentiment analysis is the main research direction in the field of natural language processing. Its main research focuses on word vector representation, text feature extraction, model establishment, etc. This article explains the current research status of sentiment analysis using machine learning technology in the industry, and proposes an improved ML-SVM (Machine learning support vector machines) public opinion information classification algorithm based on machine learning. A new public opinion word discovery mechanism is established through the characteristics of public opinion information and Shannon's law, which solves the problem of insufficient features of the training text set; in order to better measure the close relationship between public opinion information, a public opinion text information update algorithm is established, and through different data sets Simulation experiments and comparative analysis verify the effectiveness of the algorithm in this paper.

Group activity analysis holds diverse applications across various domains, encompassing daily behavior scrutiny, sports event investigation, and social behavior analysis. What sets the methodology proposed in this paper apart in group behavior analysis is its dual focus: not only does it scrutinize the behavioral trends exhibited by the collective but also delves into the nuanced variations in individual behaviors. While comprehensively examining group behavior, this approach offers researchers a fresh lens to discern individual behavioral traits, crucial for delving into and elucidating the intricacies of group behavior comprehensively. To validate the efficacy of this novel approach, we meticulously conducted experiments using the volleyball dataset and devised a corresponding system framework. The experimental outcomes yielded satisfactory results, showcasing the potential of our model in analyzing both group behavior and individual disparities, thus paving a novel pathway for practical applications.

Offline reinforcement learning represents a pivotal area of advancement within the broader realm of reinforcement learning. Its central objective is to train an agent exclusively using behavioral data, eliminating any need for online interaction. However, relying solely on insights from offline datasets can often lead to ineffective solutions, primarily due to the mismatching between the learned policy's understanding and the actual underlying environment. Recent research efforts have tended to approach this challenge with an overly pessimistic mindset, potentially compromising the agent's robustness when encountering unseen states. We introduce a self-supervised framework tailored to mitigate this issue. Drawing inspiration from contrastive-based techniques in self-supervised learning, we treat original data as positive samples and generate synthetic data from highly uncertain regions as negative samples. To simulate these regions, we employ modified Generative Adversarial Networks (GANs) to produce samples that mirroring the distribution of previous experiences and introducing a significant degree of uncertainty in terms of behavioral policy at the same time. To bolster the policy's robustness, we impose penalties for overconfident behavior when dealing with negative data. Comprehensive experiments conducted on multiple public offline reinforcement learning benchmarks have highlighted the practicality and efficacy of our framework.

In this study, we conducted a comparative analysis of three deep learning models—CNNs (90% accuracy), LSTMs (92% accuracy), and RNNs (95% accuracy)—for skeleton-based action recognition. The research focused on evaluating each model's ability to accurately classify a variety of human actions using a Kaggle dataset. Our findings revealed CNNs' strength in spatial recognition, LSTMs' proficiency in temporal dynamics, and RNNs' overall superiority in sequential processing. The study emphasizes the importance of model selection in action recognition tasks and lays groundwork for future exploration into hybrid models.

License plate recognition technology is an important part of intelligent transportation and has been widely used in traffic detection, parking charging and other scenarios. In road video inspection applications, the captured license plates are sometimes partially obscured by roadside obstacles such as trees, street light poles, bicycles, etc. The existing license plate recognition algorithms have low recognition accuracy when blocked. To address this problem, this paper proposes a license plate recognition algorithm based on multi-frame information fusion. When the license plate is partially occluded, a single frame often cannot identify the complete license plate number. By fusing context-related multi-frame information in the video, the missing license plate information can be completed. The algorithm introduces the Non-local attention mechanism to generate a transformation matrix. The transformation matrix and the corresponding pixel dot product operation are performed to obtain a weighted and adjusted feature map. The adjusted feature map is spliced with the original input feature map in the memory module, and finally the output feature map is Character classification to obtain license plate recognition results. Experimental results show that the algorithm proposed in this paper can fully integrate the characteristic information of license plate images and can effectively improve the accuracy of partially occluded license plate recognition.

Research into early data collection is often neglected, despite the growing focus on micro hyperspectral imaging technology in the medical field. Unfortunately, the data obtained from a single micro hyperspectral image is often limited in scope, making it hard to extract features and analyze spectral data of healthy cells. To address the above issues, a micro hyperspectral image stitching method based on RGB image layer feature localization and stitching is proposed, Extract the collected data into three RGB bands, perform feature extraction on them, calculate the horizontal offset, and complete the matching and stitching of microscopic hyperspectral images based on the coordinate position relationship.By employing the mean filtering technique, data fusion can be accomplished, thus allowing for the stitching of hyperspectral images on the micro dimension for spectral information.

Aiming at the problem of slow and inaccurate positioning of defects based on deep learning transmission line defect recognition model, an intelligent detection method based on improved Yolov5 transmission line image recognition is proposed. Firstly, according to the relevant characteristics of the transmission line data set, relevant features are extracted to obtain sufficient sample data. Secondly, the lightweight model of convolutional neural network is constructed, and the model training is completed by using the obtained transmission line sample data. Finally, the location and classification of transmission line defects in complex background are realized through multiple transmission line training sets and test sets. The results show that the proposed algorithm has the highest detection accuracy, and the average detection accuracy can reach 98.7 %. It is a simple, effective and practical transmission line defect recognition method.

Electric power operation is an important component to ensure the continuous supply of electric power. However, electric power operation involves a wide range, complicated process, high risk factor, and serious accidents such as personal injury often occur. Although the rules and regulations of power safety production are gradually improved and the operation management mode is gradually optimized, the low degree of digitalization and intelligence still makes the power operation face a high risk of major accidents. In this paper, deep learning framework PyTorch and YOLOv3 algorithm are used to construct a digital and intelligent situation element extraction method for the feature extraction process of image data, build a high-altitude work identification model for electric power construction to learn and train the judgment of high-altitude work, detect workers and safety belts as targets, and accurately send out early warning signals. The results provide the data basis for predicting the risk situation of electric power operation.

Human pose estimation in crowded scenes has always been a challenging task in bottom-up multi-person pose estimation. To improve the accuracy of pose estimation in dense crowds, we propose an improved bottom-up human pose estimation model called H-DEKR, which is based on Disentangled Keypoint Regression for Bottom-Up Human Pose Estimation (DEKR). The model first enhances the coarse/fine-grained feature extraction abilities of the backbone (HRNet) by introducing different structures of Polarized Self-attention (PSA). Then, Pyramid Convolution (PyConv) is introduced to extract multi-scale information, alleviating the problem of uneven human scales. Results show that our model based on HRNet-W32 achieves accuracy of 67.1% on the CrowdPose dataset, which is 1.4% higher than the DEKR, respectively. Therefore, the proposed model in this paper is able to improve the accuracy of human pose estimation in dense crowds.

The application of visual object tracking SAM-TRACK is widespread. For the problem of robot positioning and navigation, a single laser sensor obtains less information. Adding a depth camera to provide visual information will enhance the robustness of the system. Design a system for visual object tracking based on laser SLAM. Identify and segment targets by clicking, drawing edges, and inputting text, and calculate the target position by comparing the segmentation results with the depth map. At the same time, track the target in real-time to control the robot to ultimately reach the target position. Experiments have shown that multiple input methods can effectively achieve segmentation; By comparing with SiamRPN, it has been proven that the system is more efficient.

Agricultural diseases and pests pose a formidable menace to crop production and the safety of agricultural products. To address the pressing need for rapid and accurate identification and detection of these agricultural threats, this study presents novel model enhancements to the MobileNetV3 network architecture specifically designed for discerning specific agricultural diseases and pests. The refined network structure capitalizes on the utilization of multi-level feature information extracted from images, thereby bolstering the precision and robustness of agricultural disease and pest recognition. The proposed methodology exhibits outstanding performance in agricultural disease and pest recognition tasks, indicating its strong potential for practical applications.

As an essential part of media content creativity, vector fonts are widely used in graphic design and embedded applications for their convenient editing transformations and high-quality display and output effects. However, designing high-quality vector Chinese fonts requires the designer's expertise, experience, creative inspiration, time, and effort. Therefore, it is of great practical significance to study efficient vector Chinese font generation methods to improve production efficiency and liberate manpower. For this reason, it is time-consuming and labor-intensive to manually create vector Chinese fonts with a uniform style. This paper proposes a hierarchical decoupling model (T-HD) based on Transformer. The goal of generating complex vector Chinese fonts from a single font style reference and content reference is realized. This paper proposes a hierarchical decoupling model (T-HD) based on Transformer.

In order to effectively obtain word-level semantic knowledge and address the issue of inaccurate extraction of salient information during feature extraction, this paper proposes a named entity recognition method that combines the RoBERTa-WWM (A Robustly Optimized BERT Pre-training Approach-Whole Word Masking) pretrained model with attention mechanisms. Firstly, the RoBERTa-WWM model is trained to obtain word-level semantic knowledge representation. This semantic representation is then sequentially inputted into a bidirectional long short-term memory (BiLSTM) network, where attention mechanisms are applied to assign weights to key information. Finally, a conditional random field (CRF) is used to obtain the globally optimal labels. Experimental results demonstrate that compared to traditional named entity algorithms that use character-level semantic knowledge, the proposed method achieves a 1.25% improvement in the F1 score. Moreover, further improvement of 2.06% in F1 score is achieved by weighting the key information, indicating the strong performance of the proposed model in Chinese named entity recognition task.

To address the issues of low search efficiency and the inability to dynamically avoid unknown obstacles in complex environments during mobile robots’ path planning, a novel approach called the F_TSEA_IDWA algorithm is introduced to tackle these challenges effectively. The F_TSEA_IDWA algorithm is a fusion of two improved algorithms, namely TSE_A* and DWA. Initially, the TSE_A* algorithm is employed to conduct global path planning, followed by the extraction of path nodes generated by the TSE_A* algorithm as intermediate target points for the enhanced DWA algorithm. Subsequently, the improved DWA algorithm is applied to perform local dynamic path planning. Experimental results reveal that the F_TSEA_IDWA algorithm not only avoids local optima and reaches the endpoint successfully, but also effectively avoids unknown obstacles and enhances the dynamic obstacle avoidance performance of mobile robots in complex environments. Thus, the proposed fusion algorithm offers a viable solution for the path planning of mobile robots.

Face recognition technology is a kind of biometric recognition technology that has developed rapidly in recent years, and the method based on deep learning is one of the most promising research directions. In this paper, the face recognition technology based on deep learning is reviewed and analyzed. Firstly, the mainstream data sets of face recognition technology are introduced. Then, two kinds of face recognition methods based on large-scale data and synthetic data are analyzed and verified by experiments. Finally, the advantages and challenges of the face recognition method based on deep learning are discussed, and some hot spots and future development directions in current research are prospected. However, face recognition technology based on deep learning has higher accuracy and robustness. At the same time, there are still problems such as unbalanced data sets, responses in complex scenarios, and data privacy protection. This paper provides important reference and guidance for researchers by systematically combing and summarizing the research progress and application in this field and then promotes the development and application of face recognition technology.

The transmission line channel environment is complex, and it is easy to have hidden dangers. Therefore, the power maintenance personnel use a large number of intelligent transportation and inspection equipment and technical means to ensure the safe operation of the power grid, thus accumulating a large number of hidden dangers data. This paper is based on this kind of historical hazard data based on large data analysis of transmission hazard line identification method, after data processing, according to different line number and tower number, through each line, using the dynamic sliding window algorithm from the minimum step iteration until the window of the largest weight and the total ratio exceeds the specified threshold, the data in the window is the line. In this paper, the corresponding model of processing method can realize the transmission line of line identification, based on the dynamic sliding window big data processing means, without data marking, data processing, the automatic process control, can accurately find out a hidden danger of high incidence of line, for intelligent inspection provide auxiliary decision, improve line security.

Aiming at the transmission defect business scenario with many components, various defect forms and uneven size distribution, this paper proposes a transmission defect identification algorithm based on the VIT pre-trained visual large model architecture and the ViTDet object detection training algorithm. Specifically, it uses the ViT-Large model as the backbone network and Cascade-rcnn as the framework of the ViTDet algorithm. Meanwhile, in order to solve the problem of small-size defect identification in transmission scenes with large field of view, the image clipping training strategy is integrated. Cut each image equally into four parts with an overlap rate of 20% during training. In the case of the similar false detection, the recognition rate of the large model has an improvement of about 5% compared to the traditional CNN model.

How to explore the continuous latent space of Generative Adversarial Networks to extract interesting levels has been a challenge in video game level generation. To take on this challenge, the Covariance Matrix Adaptation MAP-Elites (CMA-ME) algorithm focuses on finding a diverse collection of quality solutions on complex continuous domains. By combining self-adaptation techniques from Covariance Matrix Adaptation Evolution Strategy (CMA-ES) with archiving and mapping techniques, CMA-ME effectively maintains the quality diversity of solutions. In order to obtain various playable levels that are aesthetically similar to human-made ones, this paper defines a new set of behavioral characteristics and fitness functions for the CMA-ME algorithm. It also employs the CMA-ME algorithm to explore the latent space, while simultaneously utilizing the Mixed Integer Linear Programming (MILP) method to ensure the playability of the generated levels. We apply our new method to generate diverse playable levels for several 2D tile-based games: Zelda-v0, Zelda-v1, and Pacman. The experimental results show the high quality of the generated levels compared with the baselines and thus justify the effectiveness of the proposed methods.

Due to the pressure of urbanization and commercial development on the landscape of the ancient city streets, excessive development of commercial and standardized planning schemes has led to the loss of the original unique charm of the ancient city street landscape, resulting in a decrease in environmental quality. To this end, a simulation method for visual coordination planning between the ancient city street landscape and the surrounding environment has been proposed. Using GIS technology, collect data on the ancient city street landscape and surrounding environment, and use multi-objective functions to construct a visual coordination planning model for the ancient city street landscape and surrounding environment. On this basis, an adaptive ant colony algorithm is used to solve the visual coordination planning model between the ancient city street landscape and the surrounding environment, achieving visual coordination planning between the ancient city street landscape and the surrounding environment. The experimental results show that the proposed method has a good visual coordination planning effect between the ancient city street landscape and the surrounding environment, and can effectively reduce the cost of visual coordination planning between the ancient city street landscape and the surrounding environment.

With the widespread application of intelligent detection robots for power fittings, the stability control of their motion and gait has become a key issue of concern. This paper conducts research on this issue and proposes an effective control technique. Firstly, the motion characteristics of the intelligent detection robot for power fittings were analyzed, including mechanical structure, dynamic characteristics, and other aspects. Then, based on the wavelet neural network algorithm, a stability controller was designed to achieve smooth and stable motion of the robot. Finally, the effectiveness and feasibility of the proposed control technology were verified through experiments. The experimental results indicate that this control technology can improve the gait stability of the intelligent detection robot for power fittings, and enhance its work efficiency and safety.

Due to the similarity between storage materials, there are errors and difficulties in controlling efficiency when positioning them. Therefore, a research on accurate positioning methods for intelligent storage materials based on RFID technology and machine vision is proposed. Machine vision technology was introduced, and the RealSenseD435 depth camera was used to identify the target box. The weighted average method was used to process the image, and the distribution of pixel gradient fields in the image was used as a feature parameter for the storage materials. In the positioning stage, RFID technology was introduced, using the extracted pixel gradient field of the warehouse material image as RFID tags. Based on the rigorous relationship between the phase and the distance from the warehouse material, the position of the target material was determined. In the test results, the positioning error of the design method for different types of materials remained stable within 0.15m, and the time cost remained stable within 10.0m.

Clarifying that student portraits can provide personalized learning assistance for students, thereby promoting the development of personalized education. This article presents a new learning method based on K-means. This article collects and preprocesses students' learning behavior, and transforms it into a standardized dataset for users to use directly; in response to the shortcomings of the K-means algorithm in the learning process, this paper adopts a new K-means based learning method, which can effectively improve learning efficiency; on this basis, a new learning strategy was adopted and applied to cluster analysis of student behavior data. On this basis, this article applies the improved FOA (forest optimization algorithm) algorithm to the SVM (Support Vector Mac) classification model. The Gaussian kernel support vector machine classification model can be obtained, and the improved support vector machine has an accuracy of 0.78, a recall rate of 0.88, and an F1 score of 0.84, which is higher than other algorithms. This article helps school education managers better understand students, accurately predict students, and provide personalized services for students.

It is time-consuming to manually label the defects of the steel surface at the pixel-level. In this study, we aim to train a model for steel surface defect detection based on a dataset which is weakly labeled at the image-level. To achieve this, we propose a class activation map (CAM) method based on vision transformer (ViT), which fuses the attention map and the semantic map . We also introduce an object background discrimination module (OBDM) to alleviate the problem of irrelevant background activation. Experimental results show that, compared with other CAM methods, our method has achieved performance in the task of steel surface defect detection.

Deep learning-based methods have achieved significant improvement in accuracy in diagnosing lung diseases utilizing Chest X-Ray. However, their black-box nature and lack of interpretability reduce the confidence among physicians in the reliability of machine-generated decisions, which consistently limits their application in clinical practice. In this paper, we propose a novel interpretable deep learning model VProtoNet, which can produce heatmaps that display important diagnostic image features of lung diseases and reveal how the model makes decision based on them. VProtoNet generates heatmaps by comparing the features extracted by Vision Transformer with the prototypes, each of which signifies a typical part of a Chest X-ray image, learned within the model. Further, we simplify the heatmap into a single similarity score that can be used as the basis for model classification diagnosis. To verify the effectiveness of our model, we applied our method to Chest X-ray 14 dataset and achieved an accuracy of 72.35%. Also, we analyzed the feature maps generated by our model during the classification process, discovering that they indeed intuitively demonstrate the model's recognition and understanding of the diseased areas, which enables physicians to better comprehend the model's decision-making process.

The pipeline welding robot is limited by mechanization and programmed operation, and it is difficult to adjust the welding mode according to the actual situation of the pipeline. It has the disadvantages of long welding teaching cycle and complicated manual operation. The welding quality of the pipeline is generally low, and the basic stability of the production quality is difficult to guarantee. Based on the research of machine vision, three-dimensional scanning of welds is carried out to obtain three-dimensional point cloud data to achieve high-precision three-dimensional reconstruction of pipeline welds, which can simplify the operation process, reduce labor costs, improve production quality, and make welding robots adapt to the development needs of flexible manufacturing field. Therefore, a set of welding robot system is designed. In this paper, the depth map is obtained by SGBM, and the camera pose is obtained and optimized by the image pose + point cloud registration method to realize the three-dimensional reconstruction of the pipeline. Based on the semantic segmentation of pipeline weld images in complex scenes and the three-dimensional automatic measurement of pipeline reconstruction welds, welds were obtained based on the semantic segmentation model of lightweight backbone network Mobilenetv2+Unet, and the width of welds was measured by the centerline method.

This paper proposes a cross-section analysis method of ice-covered power lines for UAV blasting de-icing. Firstly, data of ice-covered power lines are collected through onboard LiDAR and cameras. Then, point cloud processing and graphic processing technologies are used to identify the contour of ice-covered power lines, calculate the radial dimensions of ice-covered power lines and the position of the wire in the ice. Finally, the minimum bounding box method is used to fit the cross-sectional shape of ice-covered power lines based on multi-angle data. Experimental results show that the error of radial dimension measurement of this method is controlled within 3%, the fitting result of the cross-sectional shape of ice-covered power lines is highly consistent with the actual ice shape, and the overlap reaches more than 0.85. The method in this paper provides a basis for the selection of explosive quantity in UAV blasting de-icing, reducing the risk of wire damage.

In order to improve the intelligent level of accurate personnel positioning system to deal with underground emergencies, this paper shows an intelligent warning mechanism based on fuzzy neural network. The fuzzy neural network is a multi-input single-output model structure model, which has four layers, input layer, fuzzification layer, fuzzy inference layer and output layer. The intelligent warning mechanism standardizes the data of sensors in the location of the personnel label, and then uses the trained fuzzy neural network to predict. Before the disaster accident, the staff is automatically reminded to evacuate the dangerous area as soon as possible through the sound and light alarm function of the personnel label. The test shows the intelligent warning mechanism can remind the underground staff to evacuate the dangerous area as soon as possible before the disaster accident. Compared with the 70% prediction accuracy of the single sensor warning mechanism, the intelligent warning mechanism using fuzzy neural network can achieve 95% prediction accuracy.

This paper addresses the technical problem of driverless vehicles making different decisions when encountering uncertain environments while driving, In terms of road condition detection, risk feedback, behavioral decision making and path planning, the use of vehicle wireless networking technology and model fusion technology improves the stability of data transmission, road detection and decision making for driverless vehicles; Using Matlab simulation, simulation experiments and data analysis, it was found that risk feedback is a good tool for behavioral decision making in driverless vehicles, Creation of favorable conditions for real-time monitoring of the driving process and improved efficiency of simulation planning, To support accurate decision making when vehicles are at risk and to prove the reliability of the experiment, Improving the instability of conventional unmanned driving and informing behavioral decisions for driverless vehicles.

With the needs of urban infrastructure work, the number of sand trucks is increasing. Monitoring the illegal behavior of sand trucks is a highly repetitive and wasteful position of manpower and material resources. This paper presents an improved model based on YOLO V5s to identify the illegal behavior of sand trucks. This paper proposes the CA attention mechanism module to improve the network's attention to channel information, and then uses the ACON-C activation function to introduce two learnable dynamic parameters to the activation function to increase the nonlinearity of the network. Finally, the MPDIoU loss function is used. After the structure is improved, the same training parameters and data sets are used for different models through comparative experiments. The results show that the improved algorithm improves the mAP@0.5 index by 6.8%, the accuracy rate P and the recall rate R by 5.3% and 8.1%, respectively, compared with the original algorithm.

Abstract At present, the application of digital twin technology in the construction stage of cross-sea Bridges is still in its infancy, and the research on related technology expansion and application scenarios is still insufficient, and systematic research results have not yet been formed. In view of these problems, this paper defines the connotation of digital twin in the construction process of cable-stayed bridge, explores the composition of key information during the construction of bridge, builds a digital twin platform during the construction period of bridge, and applies it in cross-sea cable-stayed bridge. The application results show that the platform can optimize the construction management and control means. Improve the transmission efficiency of all kinds of information in the construction process, and provide reference for the application of digital twin technology in the construction of similar Bridges.

The current low level of intelligence in the main transportation process of coal mines, traditional protection devices can only passively shut down after accidents occur. Although they prevent further losses from expanding, they cannot provide early warning and fundamentally avoid accidents. This article designs a foreign object recognition system based on machine vision and applies it to the protection of the main transportation belt conveyor. Use YOLOv5s as the deep learning model, and deploy the trained model to the edge computing module. Real time video is captured by an industrial intrinsic safety camera and transmitted to the edge computing module to identify the foreign matters in the coal flow. Finally, only the recognition results are transmitted externally. When a high level of foreign object danger is detected, an alarm signal will be sent to the collaborative controller, which will issue a shutdown command to the specific single controller. Simultaneously upload the processed real-time video and alarm information to the control platform for display. This design improves the intelligence, stability, and reliability of the protection of the current main conveyor belt conveyor, achieving preventive measures.

In recent years, multimodal composite guidance technology has received extensive research and exploration, and multimodal fusion methods based on deep learning have achieved performance far superior to traditional methods. However, in practical applications, there are still many difficulties in the fusion technology of infrared images and high-resolution range profiles. In response to the above challenges, this paper fully utilizes the characteristics of sensor data collection to conduct in-depth research on the application of position prior, cross modal attention mechanism, and domain adaptation in multimodal object detection problems. A multimodal object detection algorithm called CenterNet- PK, which introduces a position prior, is proposed to deeply fuse infrared radar feature information while alleviating data heterogeneity issues. A feature extraction module based on one-dimensional convolution and bidirectional gated cyclic unit (Bi-GRU) was designed, and a radar feature map generation algorithm based on position prior was proposed. A cascade strategy was used to deeply fuse infrared and radar feature maps. Finally, the improved adaptive key point loss function was used to complete network training. The experimental results show that the proposed algorithm has high detection accuracy and anti-interference ability.

At present, multi-agent reinforcement learning(MARL) is in the initial stage of development, and research on strongly coupled multi-agent reinforcement learning is not in-depth. We propose a deep reinforcement learning network architecture for multi-heterogeneous robot systems based on a global status simulation platform. This architecture is based on clipped PPO and constructs a strategy model by constructing the full state action space and full action space of multi-heterogeneous robots. We trained and tested the algorithm on a multi-heterogeneous robot simulation platform, and the results showed that the proposed method can effectively target artificially fixed strategies and achieve victory.

Due to interference sources such as dial cleanliness, light pollution, light angle, and severe weather during the reading process of outdoor instruments, a design of outdoor instrument cleaning and reading devices based on wireless positioning is proposed. The focus is on the influence model of reading distance on outdoor instrument reading, composition, deployment, and positioning accuracy testing of wireless positioning. The feasibility and utility of wireless positioning in outdoor instrument cleaning and reading are verified through practical applications. Outdoor instrument cleaning and reading devices can be operated independently through wireless positioning and laser rangefinders. The entire process does not require edge recognition, image rotation, translation, or deformation conversion of instrument panel images, achieving accurate reading and reliable reading of outdoor instruments. This reduces the outdoor work intensity and improves the efficiency for workers.

This paper analyzes the importance of digital technology for the aviation equipment industry chain. Flight test as a bridge connecting the development and service of equipment, the construction of digital flight test mode through business modeling plays an important role in improving the quality and efficiency of flight test, as well as the progress and quality of equipment development. BOM definition technology is used to define the business model of the whole life cycle of flight test. Based on the flight test business XBOM architecture, describe the structure, mapping relationship, and attribute information of each BOM based on the association relationship between various departments of the test flight business. Taking flight test requirement management BOM as an example, a detailed design was carried out based on a single data source. Final list three application effects of flight test business XBOM.

In the field of algorithm research, the problem of single-source shortest path has been discussed for a long time. In this paper, the problem of single source shortest path based on Q-learning is studied. Q-learning is a reinforcement learning algorithm that iterates to update the Q value of each state-action pair to determine the optimal path. When solving the single-source shortest path problem, each node in the graph is regarded as a state, and the edge between each node as an action. By using the Q-learning algorithm, we can find the optimal strategy, that is, to minimize the total cost from the starting point to the end point, and then find the shortest path. This paper introduces the implementation process of Q-learning algorithm, including initializing Q-table, selecting action, executing action, updating Q-table and so on. Finally, the effectiveness and feasibility of Q-learning algorithm are verified by experiments.

In the field of image processing, particularly in the realm of analyzing bomb blast sites, the speed and accuracy of fine-grained segmentation have been longstanding concerns. To address these challenges, this study proposes an innovative approach based on an enhanced version of the SOLOv2 framework. By integrating low-level feature information extracted from the backbone network into the prediction network, and replacing top-level feature extraction with sub-top-level information, both the segmentation speed and accuracy of bomb blast sites are significantly improved. The experimental results demonstrate that the enhanced SOLOv2 achieves a segmentation speed of 20.35 frames per second (FPS) for small-scale blast sites, with a 3.21 improvement in precision. Notably, the average accuracy of blast site segmentation reaches an impressive 39.56. The segmentation test results robustly indicate that the enhanced SOLOv2 not only ensures efficient segmentation of small-scale bomb blast sites, but also enhances overall segmentation accuracy, thereby highlighting its superior segmentation performance.

In the future, the products tend to be small batch, multi-variety development trend, the production line can be quickly changed according to different workpieces, and reasonable production layout can minimize the workpiece change time. Therefore, from the point of view of reasonable production layout, with the path of minimizing AGV as the optimization goal, the twin checkerboard model was established, and the digital lattice selection strategy and step size algorithm were used to optimize the station layout. Finally, the optimization of the production layout of the two processes was verified at the Tecnomatix Plant Simulation with five stations of different sizes and shapes. The results show that compared with the two random layout schemes, the AGV motion path is more than twice shorter, which is of great research value in the production exchange of parts with similar structure.

In order to solve the problems in urban traffic flow and congestion management, this paper established a comprehensive management model based on the ideas of traffic management and optimization, by determining the indicators of traffic flow, congestion degree, traffic efficiency, and so on. The model was implemented and validated through the following steps. First,this paper provided a traffic flow prediction method based on the LSTM model. By training the historical traffic data, the model could accurately predict the future traffic flow. In addition,we used the TOPSIS comprehensive evaluation-Dijkstra algorithm for route planning optimization. Finally,by optimizing signal control, traffic flow control, and route planning, the effectiveness of the model was evaluated and an optimal path algorithm was provided to drivers to minimize their travel time. We concluded the model as the Intelligent Transportation System framework. Our model has good applicability in practical applications and can reasonably solve urban traffic flow and congestion problems. The model not only fits the actual situation, but also features strong practicality and high efficiency of the algorithm, providing an effective support tool for the urban traffic management department.

This article discusses the importance of indoor positioning technology and the challenges it faces, with a particular focus on fingerprint positioning technology based on Channel State Information (CSI). It delves into the rich channel feature information provided by CSI and its applications in indoor environments. However, over time or due to variations in the signal itself, indoor localization deployments may experience a decrease in accuracy or even become ineffective based on the initially collected data. This paper addresses the issue of decreasing model localization effectiveness due to temporal or environmental factors using knowledge from the domain adaptation area of deep learning. The aim of this paper is to minimize the domain discrepancy by reducing the distance in the feature space between the source and target domains, as well as the distance in the semantic feature centroid. As a result, it achieves an improvement of 32.70% and 33.59% over the performance degradation of the original model on the two target domains, respectively. When using only unlabeled data from the target domain, this approach outperforms three other basic deep domain adaptation methods.

Recently, there has been widespread interest in arbitrary-shaped scene text spotting, showing notable progress in both speed and accuracy. However, common issues such as model complexity, slow inference speed, and challenges in deployment persist. For this reason, this paper proposes Weakly Supervised Point-collection Network (WSPNet) for arbitrarily shaped scene text, which adopts deformable convolutional single-shot network architecture to realize multi-branch parallel reading of text, and combines weakly supervised text line point sampling to realize sampling recognition and transcription of arbitrarily shaped text instances. To enhance the robustness of multi-scale fusion, we utilize the Text Feature Fusion (TFF) module to dynamically acquire key features at different scales during the fusion process. This approach aims to reduce discrepancies in scale fusion. Extensive experiments on arbitrarily shaped benchmarks show that WSPNet achieves competitive accuracy and speed. For example, the proposed method achieves an end-to-end text recognition F-measure of 63.4 on the Total-Text dataset at 35.5 FPS, which is competitive with the best results in terms of speed and accuracy balance.

With the development of information technology, most domestic companies mainly use Internet recruitment methods for talent recruitment. Various Internet platforms have a large number of talent resumes. How to quickly and accurately discover and recruit talents that meet the company's development needs has become the focus of many companies and scholars. As the number of job-seeking users continues to rise, the number of resume resources increases dramatically, and a large number of resumes are misclassified. Therefore, in the massive dynamic resume resources, correct classification faces great challenges and is an urgent problem to be solved. Aiming at the low efficiency of manual resume screening in human resources recruitment, this article studies building a job talent model before screening resumes, obtaining key resume data on the Internet platform, and combining it with machine learning algorithms to provide a solution for automatic resume screening. This article proposes a dynamic resume resource classification and screening method based on machine learning, and a resume screening method that is dynamically adjusted based on changes in corporate talent requirements. Experimental results show that this method can effectively reduce the company's human recruitment costs and improve the efficiency of talent recruitment.

In order to solve the problem that the bearing vibration signal identification under a single working condition is difficult to be applied to the fault identification and diagnosis of rolling bearing in the actual production, this paper proposes a classification method of bearing fault mode based on mixed working condition. In the actual industrial production, bearings are often in a variety of complex working states. The bearing fault diagnosis research under mixed working conditions can more accurately restore the working conditions in real industrial scenarios, and improve the practicability and reliability of fault diagnosis. First of all, the timing is realized through the gram angle field conversion, and the time dependence and correlation of the original time series of bearing vibration are retained, which realizes the rich representation of bearing vibration signal and efficient fault type identification. Secondly, in order to improve the fault identification rate, the advantages of EfficientNet and Vision Transformer (ViT) models in representation learning and feature extraction are given full play through the cross-domain fusion of EfficientNet and ViT models.Comparing the experiment of various classification methods show that the method proposed in this paper can effectively extract the fault characteristics of vibration signal, identify the correct rate is significantly higher than the typical classification method, in the migration experiment of two data sets, the classification method of mixed working condition has achieved good migration effect, indicating that the fault mode recognition method of mixed working condition is easier to be applied in the actual production.

The reliability of power grids is of paramount importance to modern infrastructure. Precise fault location estimation is crucial for the efficient operation and rapid recovery of electrical networks following outages. This paper presents a novel approach using Graph Attention Networks (GAT) to improve fault location estimation within power grids. Leveraging two years of real-world data from a power grid's monitoring system, encompassing 200 fault instances, our model demonstrates a significant advancement over traditional methods. The GAT model capitalizes on an attention-driven mechanism, providing a dynamic and focused analysis of the grid's topological data, which enhances the accuracy of fault detection. Comparative experiments show that GAT model outperforms benchmark algorithms, Graph Convolutional Networks (GCN), and Graph Neural Networks (GNN), with lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics. The results highlight the GAT's potential as a robust and reliable tool for fault diagnosis in power grids, promising substantial improvements in operational resilience and maintenance efficiency.

Under the background of "double carbon", how to reduce carbon emissions is widely concerned by all walks of life. The purpose of the project is to predict and analyze the contribution of the development of new energy vehicles in China to the reduction of traffic carbon emissions by using the data provided by official organizations such as the National Bureau of Statistics and data on new energy vehicles. The grey forecast method is used to predict the quarterly increase of each factor. The following two machine learning techniques (support vector machine and ridge regression) were used. Analyze and compare the accuracy of the two models to the prediction results, and select the most accurate model for analysis. The results show that the development of new energy vehicles plays a crucial role in reducing carbon emissions.

The Natural Language Processing has achieved great success theses years, especially in sentiment analysis, text generation, and Human-Machine Interaction. Since the chat generative pretrained transformer developed by OpenAI in November 2022, more chatbots based on Large Language Models have appeared in our life, bring us a more convenient service. This paper we developed a organization self-use chatbot, using the Markov Model and the Term Frequency-Inverse Document Frequency algorithm. We use Term Frequency-Inverse Document Frequency to vectorize user’s question and calculate the text similarity to match answer, use the Markov Model to predict user’s intentions when the question is too short. This kind of chatbot is suitable for some organizations, for it can be trained with less corpus and topics, and it also has a good performance on user guidance and user satisfaction.

In the age of data and information, the amount of data is increasing and the privacy protection of data is becoming more and more important. Federated learning model is a distributed deep learning model that can aggregate model parameters from multiple clients while protecting private data. For the scenario of an industrial site where each factory cannot share its private data, the federated learning model allows each local factory to train its own model based on its own data, and then upload the model to a central server for continuous iteration and training. However, because the quantity and quality of data varies from one client to another, this can have an impact on the accuracy of the model. In order to reduce the above impact, an optimisation method that adds a weighted adjustment processing to the model on top of the federated learning model (FedAvg) is proposed. The final experimental results illustrate that the model obtained from the training of this optimisation method has better accuracy compared to the traditional federated learning model.

The measurement and extraction of relevant information among variables belongs to the research scope of information theory. In the index evaluation system, there is a common correlation among index variables. The weighted summation method is generally used for the evaluation of indexes, and the relevant information among indexes will be repeatedly calculated, resulting in a large evaluation result. Therefore, this paper proposes a calculation optimization method for measuring and extracting relevant information between index variables. By studying the index correlation types, this paper defines the staple existence forms of relevant information, uses the more applicable global sensitivity algorithm to estimate the amount of relevant information, combing with the weight method to extract and reduce the amount of relevant information from relatively unimportant indexes, to realize the correction of index evaluation results. Based on the university journals cited statistics to verify the method mentioned in this paper, the results show that this method can effectively correct the index evaluation results.

Internet of Things (IoT) technology is widely used across various domains, generating a substantial volume of data from IoT devices. These data often contains sensitive information pertaining to individuals and organizations. As a result, ensuring the privacy and security of IoT data has become a top priority. Many users prefer to connect their IoT devices to semi-trusted IoT cloud platforms. However, existing IoT cloud platforms often fail to adequately safeguard data privacy. Furthermore, most IoT data privacy protection schemes tend to overlook privacy and security concerns within this context. To address these challenges, this paper introduces a solution for IoT data privacy protection. This solution builds upon the existing architecture of IoT cloud platforms and incorporates blockchain and proxy re-encryption technology to enhance the privacy and security of IoT data. This approach achieves decentralized access control through blockchain technology and employs proxy re-encryption for end-to-end encrypted data transmission. We have developed a prototype system for this solution, including IoT client and server applications, Distributed Applications (DApps), and blockchain smart contracts. Experimental results demonstrate that this solution delivers excellent encryption and decryption performance, thereby ensuring data privacy and security. In comparison to existing solutions employing Attribute-Based Encryption (ABE), our approach offers enhanced feasibility and effectiveness. It presents an improved path for existing IoT cloud platforms to securely and reliably handle sensitive data, with promising applications across various domains.

Drone is widely applied to human production life with the development of drone technology. Therefore, it is very advantageous to carry out research on target detection in UAV images. Using target detection technology, achieve accurate capture of the target, thereby achieving accurate positioning of the target, and achieving the goal of meeting various mission requirements. Unmanned aerial vehicle (UAV) images have the characteristics of large changes in target scale, small overall target size, and are limited by viewing angle, which pose a great challenge to the detection of UAV targets. This thesis mainly focuses on the research of target detection in UAV images. This article proposes an optimization method for small object detection based on Faster R-CNN technology to address the problems of low accuracy in small object detection and missed detection of multi-scale targets in existing drone image object detection algorithms. The improved feature pyramid network structure FPN adopts top-down, bottom-up, and horizontal connection structures. Reduce prediction box bias caused by pooling regions of interest through feature pooling using bilinear interpolation. Finally, the improved algorithm was experimentally validated on the MSCOCO dataset. The faster and Faster R-CNN small object detection algorithm proposed in this article improves the average accuracy of the optimized algorithm from 91.16% to 96.13%. The Faster R-CNN algorithm has improved by 4.97% compared to the original one. The speed of the improved algorithm has been increased to 37FPS, which is approximately doubled.

Decomposition-based multi-objective evolutionary algorithm can achieve good convergence and diversity performance in solving many multi-objective optimisation problems, but it is difficult to ensure good diversity of pareto optimal solutions in some multi-objective optimisation problems with complex pareto fronts, such as fronts with large curvature or discontinue fronts. If pareto solutions are not uniformly distributed on the front surface, and there are many invalid reference vectors. To solve such problems, a multi-objective evolutionary algorithm based on reference vector interpolation is proposed in this paper. In the proposed method, the MOEAD algorithm is firstly used to obtain the initial pareto frontier, and the obtained pareto solution is clustered to obtain multiple clustering subsets. Then, is a nonlinear fitting is performed in the objective space to construct a pseudo-pareto frontier to approximate the subsets, and a new weight vector is generated by using uniform interpolation on the pseudo-pareto frontier. Finally, the new weight vector is used to update the current reference vector to push the population to move to the optimal pareto frontier and promote the convergence of the evolutionary process. In order to verify the effectiveness and advantage, the proposed algorithm is compared with nine existing multi-objective evolutionary algorithms on 17 test problems. The results indicate that the proposed algorithm can achieve better diversity of pareto-optimal solutions on most test problems.

In response to sudden water pollution incidents in rivers and canals, a pollution source tracing algorithm is proposed, employing the Markov Chain Monte Carlo (MCMC) method for rapid identification. This algorithm converts the traceability problem into a Bayesian estimation issue and utilizes the Metropolis-Hastings (M-H) sampling algorithm to sample the posterior probability density function. Consequently, it provides probability distributions for the location, time, and mass of pollutants in river canals.

This article presents and implements a firmware upgrade solution for embedded systems, which aims to address the vulnerability of using the same file that can be easily copied or stolen. When upgrading firmware, a txt file is sent by the upper-level software, which is encrypted using the SM4 algorithm. This encryption is based on a hardware-unique ID and a pre-stored upper-level software-unique ID, which act as encryption keys. During the upgrade, a secure downloader utilizes its hardware-unique ID and a pre-stored upper-level software-unique ID as decryption keys to decrypt the file. If the decryption key does not match the encryption key, the decrypted txt file will not be readable. Once decrypted, the firmware is downloaded to the target device using BSL (Bootstrap Loader) technology. Practical tests have shown that this method can achieve "one file, one key," effectively preventing unauthorized use of the upgraded firmware.