This PDF file contains the front matter associated with SPIE Proceedings Volume 13167, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Precipitation changes associated with tropical cyclones (TCs) have received increased attention, revealing distinct trends in convective and stratiform rain rates within TCs in recent decades based on satellite observational data. However, given the limited contribution of TCs to total precipitation, a notable knowledge gap remains regarding the evolution of convective and stratiform rain rates in non-TC conditions over the same period. Employing 22 years of precipitation data from the Tropical Rainfall Measurement Mission (TRMM), we investigated the temporal variations of stratiform and convective rain rates in non-TC environments during TC seasons from 1998 to 2019. Our findings reveal that the observed intensification of stratiform rain rates and weakening of convective rain rates in TCs are mirrored in non-TC rainfall over all ocean basins. These insights indicate the value of precipitation products from satellite remote sensing technology and contribute to an enhanced understanding of the hydrological cycle.

With the rapid development of computer vision technology, the application field of target detection is becoming higher and higher. With the continuous upgrading of UAV technology, the acquisition of remote sensing images has become more and more simple, the spatial resolution of remote sensing images has become higher and higher, and the information of images has become more abundant. There are many small targets in remote sensing images, the target size is quite different, and the background information is complex. In view of these problems, this paper proposes an improved YOLOv5 remote sensing image target detection algorithm based on YOLOv5 algorithm. Firstly, the backbone network of YOLOv5 is replaced by Swin Transformer, and the hierarchical feature map is constructed by using the displacement window, which effectively adapts to the computer vision task. Secondly, the SA (Shuffle Attention) attention mechanism is added to the network. The experimental results show that for the public DIOR dataset, the improved algorithm improves the detection accuracy by 1.7 % while maintaining the same detection speed as the original algorithm.

In order to efficiently and quickly monitor the land damage and control of green mines, this paper carried out remote sensing interpretation of green mines on the basis of the completion of remote sensing data processing, obtained the distribution and area of land damage and restoration and control of mining areas in 2018 and 2020, and analyzed its changes. The results show that with the continuous mining, the area of land damage in mining area increases, and the increase is mainly in stope and dump; The area of restoration and control also increased, and the object of control was mainly the waste dump.

Digital Elevation Models (DEMs) depict the configuration of the Earth surface, which is essential for remote sensing image ortho-rectification. Nowadays, the ALOS DEM, ASTER GDEM and SRTM DEM are the most commonly used models, produced based on different remote sensing datasets. Geolocation errors may be introduced with different producing methods and datasets. Therefore, multi-sources remote sensing images covering various terrain in China are experimented for the analysis of ortho-rectification accuracy using different DEMs. Based on the rational function model, the geolocation error of original imagery can be reduced greatly with the grounds control points, and different DEMs are used for the ortho-rectification of different imagery. The accuracy is experimented and verified using high-precision reference images. Experimental results show that ortho-rectificated images with the ALOS DEM can achieve better accuracy compared with the SRTM DEM and ASTER GDEM, which is significant for geometric processing of remote sensing images.

Using the data from GNSS-PWV, FY-2G Satellite detection products, wind profile radar and conventional observation data, an extreme rainstorm process on 7th August 2018 in Shenyang was analyzed. The results show that PWV values decreased first and then increased 2-3 hours before the start of heavy rainfall. There is a consistent correspondence between the decrease in TBB Value and the enhancement of rainfall intensity, but the center of high precipitation is slightly westward compared to the center of low TBB Value. There are two peaks for vertical speed and signal to noise ratio (SNR) during the main period of rainstorm. The first peak time is corresponding to the peak time of the storm relative helicity (SRH) and refractive index structure constant, it shows that the storm system is under favorable environmental conditions. Peak rain intensity appears within 1~2 h after the first peak time. The second peak time is corresponding to the time of cold air intrusion. The refractive index structure constant and storm relative helicity can forecast strong precipitation.

With the rapid development of the image object detection technology, image small object detection technology has become one of the hot researches in this field. In recent years, detection transformers (DETR) technology has been gradually applied in object detection research after being proposed by the Facebook artificial intelligence research team. Since small objects have characteristics such as small size and limited information, in order to enhance the detection ability of image small object, this paper proposes the super-resolution enhanced DETR(SRE-DETR) network based on image super-resolution. Based on DETR algorithm model, image super-resolution technology is introduced. And using the characteristics of dense layout, this article design and construct the image super-resolution network structure with residuals to enhance the image quality of small objects, so as to effectively improve the feature information of image small objects and enhance the accuracy of image small objects detection. Through the experiment of this study, it is shown that whether the image after super-resolution amplification and the image of the same size after super-division processing, the detection accuracy of small objects in the image has been significantly improved.

Currently, remote sensing image object detection is faced with two major challenges: the rich proportion and high error and leakage detection rates of small-scale objects in complex backgrounds, and limited resource power in on-orbit high-performance computation. To address these challenges and improve the situation, a modification method for remote sensing image object detection, named FGBE-YOLOv5s, is proposed based on YOLOv5s, aiming to improve the accuracy of small object detection and lightweight the model parameters. Firstly, FasterNet, based on partial convolution with fewer parameters, is introduced into the backbone feature extraction network to reduce computational redundancy without reducing the spatial feature extraction capability. Secondly, in the intermediate feature extraction network, lightweight convolution and an embedded multi-head self-attention mechanism are respectively utilized to replace corresponding modules in the original network architecture and process the spliced feature maps, to reduce the error and miss detection rate and improve the accuracy of object detection. Additionally, the calculation of the intersection over union (IoU) in object detection boxes is replaced with the superior EIOU function. The improved model is experimented on open-source remote sensing datasets nwpu-vhr10, RSOD, and DOTA, and compared with YOLOv5s in terms of accuracy and model volume. The accuracy is improved by 3.7%, 1.9%, and 3.9% while achieving a 30% reduction in model volume. Lastly, the improved model is deployed to rk3568 for hardware deployment and testing. The proposed FGBE-YOLOv5s integrates multiple optimization strategies, achieving significant improvements in accuracy on remote sensing datasets and demonstrating outstanding performance through hardware validation in resource-constrained environments. This strengthens the feasibility of lightweight remote-sensing object detection methods in practical engineering applications. This object detection method can be used in multiple scenarios, including quality inspection, scene understanding, and so on.

Map image is widely used in our work and life. In order to meet the management and access requirements of massive tiles and improve the service capability of map image, this paper analyzes the characteristics of Web Map Tile Services, analyzes the storage mode of massive tiles, studies the GeoWebCache mechanism, and proposes a tile management and service technology implementation based on PostgreSQL storage. A technology system for the management and service of massive tiles is constructed.

As the key payload on the satellite platform, the space optical remote sensor often needs to be customized according to the purpose of the positioning of each satellite, which leads to the difficulty, long cycle, and high cost of product development. To find an adaptive integration technology for multi-satellite platform space optical remote sensor is the key to realize rapid satellite networking and accelerating commercial deployment. As a new revolutionary technology, information integration technology provides strong support for realizing the goal of adapting space optical remote sensors to multi-satellite platforms. Different from the traditional space optical remote sensor design, the integrated design of information system can quickly realize the product configuration, greatly improve the design efficiency of the space remote sensor, reduce cost and increase efficiency, and adapt to the development trend of the current space era. In this paper, the information integrated design technology of a space optical remote sensor for multi-satellite platform is studied, and the practical application is carried out, and good practical results are obtained. Experimental results show that the 18.75%, 20%, 30%, 40%, 5% and 5% improvements in efficiency, adaptability, cost Reduction, integration complexity, database accuracy, and reliability respectively compared to the state-of-the-art approaches.

In this paper, a body contour segmentation algorithm for assisting radiotherapy patient posing is investigated. The image data of radiotherapy patients with body parts are captured using an RGB-D camera, and the U²-Net network model is improved to segment the body contour of radiotherapy patients wearing thermoplastic membranes and fixed stents by adding CBAM and CA attention mechanisms. The body contours acquired in the CT simulation localization room are displayed in the computer equipment in the radiotherapy room, thus helping the radiotherapist to restore the real-time posture of the patient in the radiotherapy room approximately to the same position in the CT localization room. The results show that the evaluation indexes of this segmentation algorithm can reach 98.7% for PA, 95.8% for IOU, and 0.969 for DSC, which can achieve a better segmentation effect, and the extracted patient contour is more accurate, smooth, and less noisy. Therefore, the algorithm can help radiotherapists to accomplish the task of assisted posing.

Magnetic test is one of the important environmental tests for satellites. In this paper, on the basis of the existing design method of magnetic test facilities, an improved design of satellite magnetic test facilities is proposed by using such techniques as the optimization of the magnetic field simulation coil structure based on the penalty function, the linkage acquisition of the rotary angle of the turntable and the magnetic field, and the parallelism calibration based on the conversion of optical prisms. This improved design has been applied to the construction of two new magnetic test facilities and achieved the design objectives, which can provide a reference for the improvement of other magnetic test facilities.

Underwater target detection holds significant importance in domains such as marine exploration, underwater resource management, and marine scientific research. However, due to the complexity of aquatic environments and the presence of noise interference, achieving accurate detection and recognition of underwater targets has remained a challenging task. This paper presents an algorithm grounded in mathematical morphology theory for preprocessing and feature extraction of side-scan sonar images, aimed at underwater target detection and recognition. The algorithm employs operations such as erosion, dilation, opening, and closing, which are fundamental in morphological operations. These operations enhance target outlines, edges, and spatial connectivity within the images, thereby highlighting the shape characteristics of underwater targets. Subsequently, appropriate thresholding and connected component analysis are utilized to accomplish target segmentation and localization. To validate the efficacy of the proposed algorithm, experiments were conducted on a public dataset of side-scan sonar images. Experimental results demonstrate the feasibility of the introduced mathematical morphology-based algorithm in underwater side-scan sonar target detection. In comparison with methods based on deep learning, the proposed algorithm exhibits comparable detection performance with a notable efficiency advantage under constrained computational resources.

In this paper, we explore strategies to address the issue of class imbalance in the domain of semantic segmentation of remote sensing images. Addressing the limitations of traditional CNNs when processing class-imbalanced datasets, we propose a novel feature extraction architecture based on Transformer. This architecture not only compensates for the deficiencies of Transformer in processing high-resolution remote sensing images but also enhances the model's performance on imbalanced datasets. Our method achieved a significant performance improvement in high-resolution image segmentation tasks with class imbalance, realizing a 68.35% mIoU, a notable advancement over existing baseline methods. The primary contribution of this study lies in providing an effective approach to handle the class imbalance problem in semantic segmentation of remote sensing images, paving a new path for research and practical applications in this field.

In order to solve the problems of low algorithmic recall and severely missed detection in the object detection process of high-resolution and large-field-of-view remote sensing aircraft images, this paper designs a Remote sensing rotation detector (R2ODet). Firstly, this paper introduces the parameter calibration of oriented RPN and FRM to design a multiscale step-by-step detection refining decoder (DRD) to replace the R-CNN so that R2ODet can detect rotated aircraft targets and, at the same time, improves the algorithm's detection accuracy for tiny aircraft targets. Secondly, the global feature extraction module Attention Module-ResNet (AM-ResNet) is designed in this paper, which significantly improves the detection accuracy of the model. The experimental results show that the R2ODet designed in this paper improves by 4.54% and 0.55% compared with the mAP0.5 of R3Det and G-Rep, which can be applied in remote sensing aircraft object detection.

Remote sensing target detection is a component of the computer vision field, with broad applications in natural disaster early warning and military operations. This study focuses on improving YOLOv8 network structure and proposes a small target detection network, YOLOv8-RSG, which incorporates multi-scale feature fusion. The dataset is augmented using the Mosaic data augmentation method to enhance small target detection capabilities. To address the issue of suboptimal detection performance in densely populated small target areas in remote sensing images, an incentive weight factor is introduced into the confidence loss function to optimize the function. Additionally, WIOU is utilized as the bounding box loss function to improve regression accuracy. In this experiment, YOLOv8-RSG demonstrates an average precision improvement of 5.5% and 6.0% in mAP0.5 and mAP0.5:0.95, respectively, compared to YOLOv8-base. The proposed YOLOv8-RSG model effectively achieves small target detection, providing theoretical and technical support for visual research in remote sensing scenarios.

Remote sensing building image segmentation faces complex building boundary and shape challenges due to the underutilisation of local features and global geometric information, as well as the dependence on initial contours. To overcome these problems, this paper proposes an end-to-end modelling framework called ACM-ResNet, which integrates the levelset Active Contour Model (ACM) and the ResNet deep learning network. ACM-ResNet utilises evolving level-set functions to progressively optimise the accuracy and smoothness of the contours, thus effectively addressing complex building boundary and shape challenges. Most importantly, ACM-ResNet no longer relies on manually labelling or setting the initial contour, but adaptively adjusts the contour position and shape starting from the predicted initial contour, improving the automation and robustness of the segmentation process. We evaluate our segmentation model on the Vaihingen and Bing Huts datasets, and the experimental results show that our method outperforms competing methods in terms of performance.

Aiming at the problems of insufficient data volume and scarcity of target images in the target detection task of sonar images, this paper proposes an image style migration from optical to acoustic based on an improved cyclegan method based on the imaging principle and process of side-scan sonar images to realize the acoustic database augmentation so as to improve the status quo of the scarce number of samples of sonar images. In this paper, we enhance the feature extraction capability of the model network by introducing the cbam attention module to the original cyclegan generator, and evaluate the migration effect using the image quality assessment metrics is and fid. The experimental results conducted with the traditional method show that the method introducing the cbam attention mechanism module has a better style migration effect for sonar image generation.

In order to retrieve sea water transparency in China's offshore waters from satellite remote sensing and realize large-scale measurements in time and space, a remote sensing inversion model of sea water transparency was established based on semi-analytical algorithm. According to the sea water transparency model, using Aqua-MODIS and Terra-MODIS satellite's daily grid remote sensing reflectance Rrs data. The data are pre-processed, and the quality control of remote sensing reflectance data and the cross-calibration of Aqua and Terra satellite data are completed. According to QAA semi-analysis algorithm and Doron algorithm, the study of sea water transparency parameters is completed, and the sea water transparency model is established, and the specific sea is made. The transparency products of sea water in this area are compared with the measured values of sea water. The model input remote sensing reflectance data of 443, 488, 555 and 667nm bands, and data level is L3m. The spatial resolution is 4km. The relative error between the inversion data and the measured data of sea water transparency is 21.6%. The results show that the inversion technology of sea water transparency based on semi-analytical algorithm is feasible, which provides important support for data verification and the measurement research of underwater sea water transparency measurement device.

Obtaining the geographical location of images through image geo-localization technology is a highly significant task. However, existing image geo-localization methods struggle with accuracy under difficult conditions such as viewpoint changes, illumination variations, seasonal changes, and occlusions. To address these challenges, we proposed an image geo-localization architecture based on a foundation vision model and feature mixing. The architecture involves truncating and fine-tuning the foundation vision model DINOv2 to extract robust image features, which are then aggregated using an MLP-Mixer-based mix module to obtain robust and generalized image global features. This architecture significantly improves the accuracy of image geo-localization under difficult conditions. Experimental results demonstrate that the proposed architecture outperforms state-of-the-art (SOTA) methods in image geo-localization accuracy. Compared to SOTA methods, our architecture achieves accuracy improvements of 6.35%, 4.06%, and 6.30% on the test sets Tokyo 24/7, Nordland, and SF-XL-testv1, respectively, with viewpoint changes, illumination changes, season changes, and occlusions.

In view of the current situation that there is less research on attitude determination in China in combination with the design characteristics of the attitude sensor carried by a certain type of satellite, and to ensure that the GF-7 satellite meets the engineering requirements of 1:10000 stereo mapping accuracy, a comparative test of attitude determination methods and the accuracy evaluation of test results are carried out in combination with the design characteristics of the attitude sensor carried by GF-7 satellite. Triad algorithm and quest algorithm are used to carry out multi star sensor data fusion test respectively; The extended Kalman filter model and the two-way filter overall weighting model are used to carry out the combined attitude determination test of star sensor and gyro, and the two-way filter overall weighting model is optimized. The experimental results show that compared with triad algorithm, quest algorithm can effectively reduce the periodic error of star sensor and provide stable star sensor fusion data for the combined attitude determination of star sensor and gyro; Compared with the two-way filter overall weighting model, the extended Kalman filter model can effectively suppress the high-frequency measurement noise of star sensor and the cumulative error of gyro drift error, and can ensure the engineering requirements of 1:10000 stereo mapping accuracy of GF-7 satellite.

Object detection is an important computer vision task that deals with detecting instances of visual objects of a certain class in digital images. It is also the basis for many other computer vision tasks, such as instance segmentation, image caption, and object tracking. Object detection has now been widely used in many real-world applications, such as autonomous driving, robot vision, and remote sensing image detection. Especially With the development of remote sensing technology, the requirement of remote sensing image object detection technology is further improved. It is of great significance for the development of remote sensing image detection to recognize the target of remote sensing image quickly and accurately. However, supervised training of traditional object detectors requires well-labeled large-scale datasets, and for the problem of low detection accuracy on remote sensing images with dense targets and poor labeling quality, a object detection method for weakly labeled remote sensing images based on adaptive weighted fusion algorithm is proposed. Object detection experiments are done on the RSOD data set and part of the DOTA data set under different degrees of localization annotation offset, and the experimental results show that the optimization method achieves a mAP value of 73.2%, which is 10.6 percentage points higher than the original algorithm, and it significantly improves the detection accuracy of inaccurately localized and annotated targets under the maximum degree of offset of the RSOD data set.

This paper aims to enhance the Oriented R-CNN framework for object detection in aerial images. Leveraging a diverse collection of high-resolution aerial imagery, including buildings, vehicles, ships, and aircraft, we propose improvements to address challenges like object scale, orientation, and occlusion common in such images. While the Oriented R-CNN method has shown promise in handling oriented objects, challenges in scale and spatial awareness remain. We introduce a novel approach that enhances the Oriented R-CNN framework by integrating dynamic heads, combining scale and spatial awareness. Through scale-aware attention mechanisms, we address issues with object scale, and spatial-aware mechanisms capture fine-grained spatial information. Extensive experiments, including the Dota dataset, validate the method's superior performance. Our contribution enhances object detection precision in remote sensing, overcoming previous limitations.

Simultaneous Localization and Mapping (SLAM) is a core technology for autonomous driving of intelligent vehicles. Based on the single-sensor mapping and localization method, it cannot tackle the issues of environmental degradation and low sampling frequency, and the accuracy and robustness of the mapping and localization solutions, which needs to be further improved. With the development of Multi-sensor Fusion methods, it is possible to make full use of the complementary performance advantages of multiple sensors to fuse the localization and mapping solutions of multiple sensors, which can better meet the requirement of high-precision localization in most scenarios and enhance the robustness of the algorithm. Therefore, this paper proposes a SLAM algorithm for mobile robots based on graph optimization. Due to the large workload of the Cartographer algorithm for adjusting parameters, an optimization method is designed, and a suitable range of parameter values is processed according to this method. With the aim of improving map building accuracy and localization stability, a detection method of Scan Context fusion history frame closed-loop is proposed to improve the closed-loop detection accuracy in the LIO-SAM algorithm. Finally, mobile carts of indoor and outdoor are established in real scenarios. The paper shows that the Cartographer parameter optimization scheme is feasible and the Scan Context closed-loop detection algorithm is more robust, which improves the autonomous localization accuracy of the mobile robot in indoor and outdoor environments.

Aiming at the visible light localization system in the fingerprint acquisition link, which is time-consuming and inefficient, this paper proposes a visible light localization system based on generative adversarial network and semi-supervised learning to realize high-efficiency fingerprint acquisition and high-precision localization. Firstly, the image of the lamp is captured using a CMOS smartphone, and the fingerprint database is constructed by feature extraction of the feature image of the lamp using image processing technology, which can increase the diversity and richness of the fingerprint database; secondly, the unlabeled pseudo-fingerprint data is generated by GAN network, which expands the dataset, thus reduces the cumbersomeness of fingerprint acquisition and improves the efficiency of fingerprint acquisition; lastly, the real fingerprint data and pseudo-fingerprint data are mixed with semi-supervised training of CNN models for location prediction. The experimental results show that in the experimental space of 2.25m×1.65m×3m, only 1564 labeled fingerprint data are collected in this paper, and the achieved average error of localization is 2.48cm. among them, the probability of error less than 5cm accounts for 90%, which meets the indoor localization requirements.

Multiple studies about the indoor positioning technologies have been reported. However, literature involving scientometric analysis of indoor positioning technologies is sparse. Here, we use scientometric analysis and a historical review to highlight recent research on indoor positioning technologies. We use the former to examine research on indoor positioning technologies from 2002 to 2022. The latter is used to identify the most frequent keywords in keyword analysis, as well as explore hotspots and indoor position technology trends. Scientometrics analyzed the mainstream positioning technology, changes in research topics, and major research trends, and found that there were 2739 research papers on indoor positioning technology, a 4-fold increase over the last 20 years. These findings offer a vigorous roadmap for further studies in this field.

The article takes the coal mining subsidence area of a mine in northern Shaanxi Province as a test site, and constructs the initial subsidence model by utilizing the aerial image data from a low-altitude UAV. The visible band difference vegetation index (VDVI) is used to eliminate the subsidence data of the vegetation covered area, which has a large influence on the accuracy of the subsidence model, from the initial subsidence model, and then combined with the remote sensing image enhancement algorithm to fit the subsidence information of the vegetation area, so as to generate the subsidence model of the ground surface with the influence of the vegetation eliminated. The modeling method of remote sensing image enhancement algorithm after removing noise and correcting the influence of systematic error can meet the basic requirements of large-scale and high-intensity coal mining subsidence monitoring in western low-vegetation-covered areas, which provides a feasible solution for the use of UAV remote sensing technology for efficient monitoring and fine modeling of coal mining subsidence in western mining areas.

In response to the significant difficulty in detecting extremely small targets in aerialremote sensing images, this research proposes a data enhancement method that optimizes the negative example selection strategy, that is, to achieve the objective by optimizing the strategy of selecting negative samples for training during the training process on the basis of existing deep learning-based target detection methods. Taking the Faster-RCNN model as an example, we use the outcome of model error identification as the negative example in the next epoch of the training process and adjusts the loss function according to the negative example category. Experimental shows that the algorithm can effectively enhance the detection accuracy of extremely small targets in remote sensing images, with a strong screening ability for interference regions in complicated backgrounds.

To reduce the impact of magnetization direction and stripe noise on the positioning accuracy of magnetic target and improve the positioning accuracy of underground shallow unexploded ordnance with different attitudes, this paper studies a normalized magnetic source intensity localization method based on Gaussian convolutional kernel. The traditional normalized magnetic source intensity method is not affected by the magnetization direction, but the stripe noise generated by the method itself can affect the positioning result. The use of Gaussian convolutional kernel can eliminate the influence of stripe noise. The actual flight test shows that the proposed method can eliminate stripe noise and is not affected by magnetization direction, and the maximum positioning error for five different attitudes of simulated unexploded ordnance on the ground is less than 15cm. Therefore, this method can be used for high-precision positioning of underground shallow unexploded ordnance with different attitudes.

For ecological environment protection research and regional sustainable development, vegetation, as a crucial land cover type, necessitates long-term, large-scale, and highly accurate monitoring. This paper formulates an enduring vegetation monitoring approach using Landsat satellite remote sensing imagery on the Google Earth Engine (GEE) cloud platform. By computing the Normalized Difference Vegetation Index (NDVI), we formulate a method for prolonged remote sensing monitoring of vegetation, unveiling its temporal and spatial pattern evolution. Firstly, based on Landsat remote satellite imagery and GEE, we calculated the NDVI. By leveraging the Normalized Difference Water Index (NDWI), we removed inland water bodies and outliers within the study area, enabling the precise generation of spatiotemporal NDVI distribution maps. Subsequently, the NDVI was analyzed for time trends in conjunction with grade transfer matrices. Finally, combined with Moran's I, high/low clusters, clustering and outlier, cold/hot spots, etc., the NDVI spatial pattern characteristics were analyzed. Using the Zhoushan Island in China as an example, the results indicated that: (1) The most significant changes were observed in the low NDVI area (0-0.2), which expanded from 28.84 km² to 67.29 km². (2) Concentrated mainly in the inland areas are the NDVI High-High clusters and the significant hot spots, while the coastal areas primarily feature NDVI Low-Low clusters and significant cold spots.

In the field of high-precision intelligent optical remote sensing satellites, in order to accurately measure the temperature and vibration changes of the core components inside the optical remote sensing system, a high-precision intelligent optical remote sensing measurement system based on distributed fiber optic measurement sensors is designed, which meets the accuracy measurement requirements under thermal vacuum environment assessment and solves problems such as blurred in orbit images, insufficient positioning accuracy, low compensation efficiency, and difficult post-processing of high-precision cameras. It promoted health monitoring and intelligent adjustment of satellite in orbit images.

Precipitable water vapor (PWV) mirrors the dynamic changes of atmospheric water vapor and can be accurately retrieved by the Global Navigation Satellite System (GNSS). Aiming at exploring the potential of the Continuously Operating Reference System (CORS) of China Southern Power Grid to contribute to meteorology, this paper first establishes a Weighted Mean Temperature (Tm) model that is more applicable in Guangzhou. Subsequently, the PWV sequence and variation characteristics during the period of heavy rainfall from June 30 to July 5, 2022, are inversed and analyzed based on the observation data of CORS in 15 reference stations around Guangzhou. The results indicate that PWV variations bear a close relationship with the actual rainfall events. The PWV increases remarkably before heavy rainfall, remains at a high level during rainfall, and gradually returns to a stable state after rainfall stops, revealing that PWV changes can serve as an indicator for heavy rainfall monitoring. Therefore, meteorological services in the CORS of Southern Power Grid can be further enhanced in the future.

Based on PS-InSAR technology, surface deformation monitoring and analysis were conducted for the Bortala Mongol Autonomous Prefecture from 2015 to 2020. Through the processing and analysis of satellite remote sensing data, we obtained high spatiotemporal resolution information regarding surface deformation. The research results indicate that the deformation rates in most areas of the Bortala Mongol Autonomous Prefecture range from -10mm/year to 10mm/year. However, there are distinct surface deformations in 8 specific areas, with a cumulative deformation of around 300mm over the course of 5 years. The deformation is attributed to factors such as soil erosion and human engineering activities. These findings not only establish a scientific foundation for monitoring and preventing geological disasters in the region but also provide valuable insights for surface deformation monitoring in similar areas.

Comprehending the mechanisms that impact Ecosystem Services (ESs) is vital for fostering sustainable development in regions. This study quantified six ESs in the Pearl River Delta region (PRD) for the years 2000 and 2020. This study selected eight factors, considering both natural and human perspectives, to assess their impact on ESs using the random forest method. The study utilized the Geographically and Temporally Weighted Regression (GTWR) model to investigate the spatiotemporal variations in factors affecting ESs. Results indicate that the spatial change of ESs differs between 2000 and 2020. In 2000, natural factors prevailed in regional ESs, but by 2020, human factors gradually took over this dominance. The spatiotemporal heterogeneity of factors influencing ESs is distinctly different, providing insights to enhance the ecological sustainability of the region.

Not all the physical evidence and traces left by field accidents are necessarily related to the occurrence of accidents and the identification of speed. Later appraisers also need to carry out the reconstruction of traffic accidents through the analysis of these evidence. Therefore, the traditional methods for the identification of speed of field road traffic accidents are generally complicated and not accurate enough. Therefore, Inter-frame Difference calculation method is more convenient, quick and accurate in the actual traffic accident case handling.

In order to suppress noise in 3D point clouds while preserving geometric features, a non-local collaborative denoising method based on weighted nuclear norm minimization is proposed. Firstly, an iterative weighted PCA algorithm is used for preliminary normal estimation, followed by bilateral filtering and normal-based PCA for refinement. The eigenvectors of the covariance matrix constructed from the neighboring points can form the local coordinate system. Then the local point cloud is projected and transformed into a height-map patch. Subsequently, to reduce the difficulty of constructing similar structures, similar height maps are searched in the global point clouds, and assembled to form the low-rank height map patch group matrix, which is considered as a descriptor of non-local similar structures. Subsequently, the weighted nuclear norm minimization is applied for low-rank matrix recovery, thereby reducing the noise of local point clouds. Finally, the denoised local point clouds are combined to form the global point cloud, realizing non-local collaborative denoising. Based on updated positions, a projection optimization algorithm is used to recover sharp features. Experiment results indicate that the proposed method effectively reduces noise while preserving the original features of the point cloud.

Phase synchronization is a necessary problem in bistatic SAR systems. The traditional phase synchronization problem is solved by the synchronization link between bistatic SAR systems. However, the synchronization link greatly increases the system complexity and periodically interrupts the SAR data acquisition process, thereby introducing new errors. This article further discusses the phase synchronization method in the echo domain under alternating bistatic mode. Changes in terrain fluctuations may cause the phase synchronization method to fail. Therefore, an improved phase synchronization method in the echo domain under alternating bistatic mode is proposed, which adopts the mode of cross transmitting signals from the two bases and extracts sub aperture phase information, The phase synchronization of bistatic SAR is achieved using a method based on the joint processing of information in the bistatic echo domain. Simulation experiments have verified the correctness of theoretical analysis and the effectiveness of the method, providing theoretical support for this method.

At present, the traditional particle size estimation of industrial raw materials mainly relies on manual measurement and observation, which has the problems of poor accuracy and low efficiency, etc. In this paper, an automatic image-based particle size estimation method for industrial raw materials is proposed, and the YOLOv7 target detection algorithm is used to classify and detect the particles. In YOLOv7, segmentation module, Biformer attention mechanism and NWD (Normalized Gaussian Wasserstein Distance) loss function calculation method are introduced to solve the problems of the original algorithm's prediction frame is large or skewed, the speed is slow when detecting the dense and small targets, and the detection accuracy is low. The problem of the original algorithm is solved. In this paper, two particle size estimation statistical methods are proposed to classify and count the particle size of industrial raw material particles. The test results show that the improved algorithm achieves 99.5% mAP on the gravel dataset, 90.8% mAP on the industrial raw material particles dataset, and 92.7% mAP on the mixed dataset, and the prediction of segmentation of a picture with hundreds of dense small targets only takes 27.2ms, which can meet the real-time and accurate industrial raw material particle size statistics in the industrial field, and it has a greater Practical value.

Since water scenes are highly susceptible to environmental factors, in the actual process of collecting and storing pictures, it is easy to cause noise pollution of image samples. This requires that the feature extraction algorithm can reduce the impact of noise on the sample data when dealing with data containing noise, i.e.That is to say, the algorithm model is required to have high robustness. The high-speed corner detection algorithm FAST, scale-invariant feature transform SIFT, etc. are traditional feature point detection methods, but the advantages in terms of computational speed and robustness are mixed. In this paper, we focus on the SuperPoint network, which has better robustness, and modify the network accodrding to the requirements of real-time and accuracy. To address the problems of gradient vanishing and gradient explosion, a residual connection structure is added between each convolutional layer and activation function. Meanwhile, to ensure the convergence speed of the model, a normalisation layer is added between the convolutional layers and the activation function. Finally, in order to improve the representation and generalisation ability of the model, the SE-Net channel attention mechanism module is added after the residual connection structure. Some of the Seaship ships, navigation beacons and other datasets are transformed into feature point datasets in the training dataset and strengthen its feature extraction capability for water scenes.Experimental analysis is conducted based on the water scenario, and the experimental results show that the detection and matching effects of feature points are improved under the guarantee of a slight increase in the computational speed, the detection effect of the number of feature points is improved by about 15.4%, the matching effect of feature points is improved by about 9.7%, the nearest-neighbour accuracy of the NN mAp is improved by about 9%, the repeatability of the Rep. is improved by about 11.4%, and the average positioning error is reduced by about 2.1%.

3D terrain reconstruction technology can produce mathematical models corresponding to real scenes, and be applied to research fields such as scientific exploration, heritage conservation and architectural design. Taking the Mount Everest as an example, this paper realizes the construction of a real terrain model in UE4 (Unreal Engine 4) by high-precision DEM (Digital Elevation Model), and explore the cause of the step-like surface of the model and solve it. Using UE4 for reconstruction can improve the efficiency of using terrain models and help terrain models into subsequent applications.

The spatiotemporal characteristics of actual evapotranspiration (ET_a) in Northeast China from 2003 to 2020 was analyzed in this research using the global SSEBOP ET_a product. Then simple correlation coefficient and sensitivity coefficient were adopted to quantify correlation relationship and the sensitivity of ET_a to the influencing factors such as air temperature (T), precipitation (P), surface net radiation (RN), surface wind speed (WS), actual vapor pressure (EA), and normalized difference vegetation index (NDVI), et al. The result showed that ET_a showed obvious seasonal characteristics, with the highest ET_a in summer. Meteorological factors and land use characteristics jointly affect regional ETa. In summer, land use types played a dominant role. The WS, EA, P, and NDVI were mainly positively correlated with ET_a. The RN and T were mainly negatively correlated with ET_a. As time goes on, the ET_a will be more sensitive to RN, P and NDVI, but be less sensitive to T, EA and WS.

Existing intersection safety evaluation methods are mostly based on traffic accident data statistics, which has shortcomings such as difficulty in data acquisition, long data collection time and long evaluation time span. In order to solve these problems, this study obtains high-definition and accurate spatio-temporal trajectory data of different traffic subjects at intersections through UAV detection technology, and combines the traffic conflict theory. Three traffic safety evaluation indexes were obtained, and five intersections in Tianjin were evaluated by using grey clustering method, and the five intersections were divided into four safety levels.

In order to understand the impact of land use change on the surface thermal environment and solve the urban heat island effect, this paper designs and implements an urban land use thermal effect analysis method based on network structure and fractal theory. First, the classification of urban land use and the inversion of urban surface temperature are completed. After that, fractal theory is innovative applied to the study of urban thermal environment, and the complex network theory is introduced into the analysis of urban land use thermal effect. In this paper, Shenyang, Liaoning Province, China, was selected as the research area, and Landsat8 data were used in 2013, 2017 and 2020. The results show that from 2013 to 2017, the higher built-up area density, the higher urban temperature. The higher forest, water and plowland density, the lower urban temperature. The correlation between land surface temperature and NDVI (Normalized Vegetation Index) is negative, while the correlation between land surface temperature and NDBI (Normalized built-up Area Index) is positive.

Study the method of underground cable laying environment survey based on wide-angle image distortion removal algorithm, understand the underground cable laying situation, and improve the security of power grid. After correcting the distortion of the wide-angle depth image of the downhole cable laying environment through the nonlinear model method, the cable nameplate coordinates are located, and the raspberry Pie carries its own high-definition camera to shoot the cable nameplate, Obtain the environmental information of cable laying space and cable nameplate, and collect and display the environmental information of temperature, humidity and gas concentration in the cable well using temperature, humidity and gas sensors. The experimental results show that this method can effectively survey the underground cable laying environment, not only can accurately capture the spatial environment information of underground cable laying, accurately identify the environmental information contained in the cable nameplate, but also can simultaneously detect the temperature, humidity and various gas concentrations in the power well, which has a good application effect.

The purpose of this paper is to explore the geographical information and spatial layout of Guiling Village, a Dong nationality village that has been declared World Cultural Heritage. Guiling Village is an administrative village located in Bajiang Town, Dong Autonomous County of Sanjiang, Guangxi. It consists of 7 natural villages: Guiling Tun, Jiangtou Tun, Liugeng Tun, Nianmatun, Guidong Tun, Guibun and Sanhe Tun. In the process of applying for World Cultural Heritage, it is necessary to collect geographical information such as geographical environment and resources of traditional villages, external spatial layout and internal spatial structure of villages. In order to improve work efficiency, we adopted UAV aerial survey technology to model traditional villages and draw topographic maps. In the process of geographic information collection, we mainly use UAV aerial survey technology. UAV aerial survey technology has the advantages of high precision, high efficiency and high resolution, and can quickly obtain the geographical information of traditional villages. Through the photos and videos taken by drones, we can obtain the geographical information of the village's terrain, buildings, roads, water systems, etc. At the same time, professional software is used for modeling and analysis, providing scientific basis for the protection and planning of traditional villages.

Shenmu City, Yulin City, Shaanxi Province, is the largest coal-producing city (county) in China, and the rapid development of the mining industry has led to the problem of large-scale regional ground subsidence, which has become a major problem for the city. According to statistics, there are more than 200 disaster hazards in the city, including collapse and landslides, but the monitoring of ground deformation in the city is still blank. In this context, this paper uses SBAS-InSAR (small baseline subset InSAR) technology to monitor and analyze the ground deformation in Shenmu City. The results show that there are three main subsidence zones in Shenmu City, and there are several discrete subsidence zones within the main subsidence zones, and the maximum cumulative subsidence in the subsidence zones reaches -105 mm/a between January 2020 and December 2021. In addition, seven typical mining areas were selected from the three major subsidence zones, and the spatial and temporal variations of subsidence were analyzed for the seven mining areas. The results can provide a basis for the municipal construction of Shenmu City and the planning of safe production in each mining area, and make up for the gap in monitoring data of ground subsidence in large areas of the region.

Precipitable Water Vapor (PWV) with high spatial and temporal resolution can be obtained by ground-based GNSS technology, but the detection accuracy is affected by the atmospheric weighted mean temperature model and the calculation accuracy of the wet delay. In this paper, an interpolation method taking into account the prior water vapor information of GNSS is proposed, and the radiosonde data is used as the prior water vapor information to establish the atmospheric weighted mean temperature model in Xuzhou. Then, the high precision PWV value is obtained by the weighted average of PWV obtained from the atmospheric weighted mean temperature model and PWV obtained by GAMIT. Based on the GNSS data of the Xuzhou area for 92 days from June to August 2019, this paper uses the new scheme, inverse distance weighting, local polynomial, radial basis function, and Kriging respectively to interpolate to PWV at the radiosonde station. The results show that the interpolation accuracy of the new scheme is the highest, RMSE is 1.9mm, and the average improvement rate of RMSE of the new scheme is 3.675% compared with the four interpolation methods.

CarMaker is a widely used simulation testing software, yet there is currently no scenario library available for standard regulations in the market. This article presents the development of a scene library conversion service that converts OpenXbased scene libraries into CarMaker scene libraries. This service has the potential to enhance the efficiency of host factory scene simulation by 95% and expand the scale of CarMaker's scene library, achieving a 100% conversion rate of scene elements. This service swiftly converts scenarios with the aid of a blend of C++ and C # programming, ensuring data integrity and accuracy. We expanded the usage scenarios of the standard scene library and resolved the issue of a limited number of scenes available in the CarMaker scene library within the simulation market.

In the field of tomato disease spot detection, despite significant advancements in convolutional neural network technology, challenges still persist in detecting small-sized tomato disease spots. This study aims to enhance the tomato disease spot detection model to improve the accuracy of detecting small disease spots on tomatoes. Firstly, the concept of DenseNet is introduced, and a Dense_C3 module is constructed, which incorporates two 3×3 convolutional layers within the Bottleneck of C3 to achieve dense feature propagation and reuse. Secondly, the Convolutional Block Attention Module (CBAM) is employed to weight the feature maps, enhancing the focus on local and spatial information to enhance model accuracy. To validate the effectiveness of the proposed improvements, experiments were conducted on tomato disease spots of different scales, including Septoria Leaf Spot, Early Blight, Late Blight, and Tomato Yellow Leaf Curl Virus. The experimental results demonstrate that the introduction of the Dense_C3 structure and CBAM improves the detection accuracy for small-scale targets. In a comprehensive evaluation, the improved YOLOv5s model shows increases of 2.4% in F1 score and 2.1% in mAP.

In response to the current issues of missed detections, false alarms, low detection accuracy, and incomplete detection categories in potato quality assessment, this paper proposes an enhanced YOLOv5s-based object detection model. The model categorizes external potato features into four standard classes. Firstly, this paper introduces the SE (Squeeze and Excitation) attention mechanism, enabling the model to adaptively adjust feature weights across different channels, emphasizing features crucial for quality assessment. Secondly, incorporate the SPPCSPC (Spatial Pyramid Pooling Cross Stage Partial Channel) convolutional neural network structure, transforming various-sized feature maps into fixed-length feature vectors. Additionally, employ the up-sampling operator CARAFE (Content-Aware Reassembly of Features) to enhance the performance of the feature pyramid network. Experimental results demonstrate that the improved YOLOv5s model exhibits a 4.5% increase in detection accuracy and a 2.1% improvement in average precision.

Previous convolutional neural networks (CNNs) in the field of semantic segmentation, particularly in road scenes, suffer from overfitting, insensitivity to positional information, and limited robustness due to convolution and pooling operations. In this paper, we propose a multi-scale multi-feature fusion self-attention network (MSMA-Net) based on the U-Net architecture. The Decoder stage of the U-Net structure is removed, retaining only the first four sampling layers of the Encoder. The final features from each layer are simultaneously fed into a multi-scale pyramid pooling structure, where different scale pooling operations merge the features into a unified dimension. The output features are then passed through a Transformer Encoder stage and the MLP Head to produce the final classification results. The proposed method is trained on the Cityscapes and CamVid datasets, with half of the data randomly selected for training. The achieved average intersection over union (mIoU) scores are 77.9% and 72.2% on Cityscapes and CamVid, respectively, demonstrating significant advantages over other networks.

Lanzhou lily is the only kind of sweet lily in China. However, its yield and quality have decreased significantly in recent years due to gray mold disease, bulb rot disease and other diseases. In order to improve the anti-interference ability of Lanzhou lily diseases diagnosis model, only 4, 8, 16, 32, 32 feature maps were selected from the VGG16 model's five pooling layers removing 75% of all feature maps. In the process of fusion, it was found that the average pooling was better than the maximum pooling and the maximum feature selection was superior to random feature selection in the analysis of the diagnosis accuracy and anti-noise performance. The result shown that the accuracy of Lanzhou lily diseases diagnosis reached 97.82% by using the multi-scale fusion CNN with the average pooling and maximum feature selection. In addition, the anti-interference ability of Multi-scale fusion CNN was obviously better than that of VGG 16 model for Gaussian noise, salt-and-pepper noise and speckle noise. The diagnosis CNN constructed in this paper can provide technical support for digitized field management of Lanzhou lily.

In order to efficiently and accurately identify and detect pavement cracks, this study proposes a pavement crack detection algorithm based on an improved Yolov5s network model. The algorithm introduces a weighted bi-directional feature pyramid network BiFPN as a neck feature network for fusing feature maps of different dimensions, thus enhancing the bottom feature information and improving the feature aggregation effect. In addition, CBAM, an attention mechanism, is employed to enhance the learning and extraction of feature information from the pavement crack image, while attenuating the influence of the pavement background, which is similar to the crack, on the detection results. The experimental results on the homemade dataset show that the improved yolov5s model improves 10%, 2.2%, and 4.6% over the original model in terms of precision, recall, and mean average precision values, respectively. This indicates that the improved algorithm is feasible for pavement crack inspection.

Aiming at the problems of overexposure and poor local contrast in traditional low-light image enhancement algorithms, an adaptive low-light image enhancement algorithm based on HSV chromaticity space is proposed, which firstly converts the input image into RGB to HSV chromaticity space, then extracts components from the converted HSV chromaticity space, and performs convolution operation on the details respectively, and then weight fusion and conversion to RGB image by the convolution operation. Secondly, the weights of the convolved components are fused and converted to RGB images, and finally, the enhanced images are obtained by histogram equalization and contrast-brightness enhancement. The experiments demonstrate that the average PSNR value of this algorithm in 11 different scenes is improved by 44.8%, 31.35%, and 17.86%. In contrast, the SSIM value is reduced by 59.63%, 56.58%, and 25.4%, respectively, compared with the contrast-brightness image enhancement algorithm, the adaptive gamma transform algorithm, and the CLAHE algorithm.

Aiming at the problem that the current laser point cloud obstacle detection calculation method has poor segmentation effect in the outdoor volcanic environment, and the time length and obstacle error detection rate are high, a partition method that can automatically evaluate the sensor height and improve the Euclidean partition method is planned. Clustering algorithm obstacle detection strategy. Firstly, the ground simulation is carried out through the segmentation method of sub-regions, and on this basis, the method of automatic recognition of the sensor height is added to solve the problem of insufficient segmentation of outdoor slopes; then, the improved version of the Euclidean aggregation algorithm is used to obtain The off-site points of the object are aggregated and framed for visualization; finally, the experimental verification is carried out on the collected user external environment data set on the real experimental platform. The experimental results show that: compared with the traditional plane simulation combined with the ground plane segmentation algorithm and the traditional fixed value Feng family algorithm, the should segmentation method and detection algorithm are ab out 16% shorter than the time cost reduction, and the positive inspection rate is about 24% higher.

Automatic line art colorization has been largely adapted to the anime industry since it can improve animation production. Reference-based method provides a more convenient and intuitive method but now is equipped with a challenge in aspects of color match and semantic consistency. In this paper, we propose an GAN-based line art colorization architecture. We introduce an attention mechanism module which fuses and merges coordinate attention and stop-gradient attention for improving encoder's ability of key region perception and feature extraction. To alleviate color overflow, we design a segmentation branch to control the region correspondence between the line art and the reference image. Multiple experiments unequivocally illustrate that our method excels over other state-of-the-art reference-based techniques in animation line art colorization task.

Cytological smears play an important role in disease diagnosis, particularly for specific cancers such as thyroid and cervical cancer, where cytological smears are primarily cell smears. With the rapid development of deep learning technology, more and more people are using deep neural networks to classify and discriminate cancers. However, getting ethically certified and clear cytological smears is hard, so training is always based on small data sets. When facing small cell smears with different levels of brightness and clarity, which are collected from various devices and environments, models' performance is often limited. Given the above situation, we have designed a cytological smear classification model that is trained on an augmented data set and considered the gray-level co-occurrence matrix of the image, making the model perform better when faced with noisy images, and we call it the Glcm-BoT model.

For prestressed concrete continuous girder bridges, a single monitoring section can be used to realize the early warning or alarm function of the overload status of the whole bridge. Obviously, the overload alarm mode with a single monitoring section can significantly improve the monitoring efficiency and reduce the monitoring cost. In this paper, combining with the internal force analysis of the The approach bridge of Yuping-Yihe grand bridge, the results of enduring condition bearing capacity limit state and regular use limit state calculation and stress calculation results, and taking into account the effects of the modal analysis of the main girder, we first solve the problem of how to monitor the overload condition of the whole bridge with the least number of measurement points or test sections, determine the critical section or control section of the main girder, and then select the monitoring section by combining with the site installation and test conditions.

Since the number of bridges in China keeps increasing and some bridges with the damage problems brought by the age of are gradually taken into account, thus the health monitoring of bridges becomes much more important. Based on the research of China’s intelligent sensing materials in the past decade, this paper summarizes the function of common sensing devices in crack pattern recognition, rust expansion monitoring, damage identification, curing shrinkage strain, and fatigue damage monitoring, and clarifies the feasibility of new monitoring technologies in practical engineering. It is also cited that the advantages and working principles of two data treatment, wavelet analysis and Long-term and short-term memory neural network, applied in the field of bridge health monitoring. The results show that intelligent sensing material and signal analysis technology has a wide application prospect in the future bridge surveillance work because of its good working performance. This paper can provide some reference for the application of intelligent sensing and signal analysis technology in the information of bridge health monitoring.

The permanent ground displacement caused by geological disasters is the main reason that induces large additional axial stress in pipelines. A combination of radar satellite remote sensing surface displacement monitoring technology and pipe body deformation monitoring technology is proposed to analyze the interaction between pipe and soil. A surface displacement measurement method based on SBAS-InSAR is proposed, and the surface displacement monitoring of millimeter-level kilometer level pipeline is realized with high precision. The finite element model of pipeline stress analysis is established by using space pipe element and nonlinear pipe-soil element considering the actual pipeline routing and process loading conditions. Taking remote sensing landslide displacement as the exact boundary condition, a method for calculating pipeline stress in landslide section is presented. The practice shows that the maximum relative error between the pipeline stress results obtained by this method and the distributed strain gauge monitoring results is less than 20%, which can effectively realize the spatial reconstruction and safety evaluation of the stress state of the pipeline in the landslide section, and provide technical basis for the establishment of the digital twin of pipeline safety state in the future geological disaster section.

This study integrates MaxEnt ecological niche model with ArcGIS technology to predict the potential suitable habitat for Xanthoceras sorbifolium Bunge in China, based on accurate and effective geographic distribution data. Three environmental factors, including climate, topography, and soil (comprising 26 environmental variables), were employed to predict the potential suitable habitats for X. sorbifolium, classify the suitable habitats, and assess its adaptability. The results indicate that the potential suitable distribution areas for X. sorbifolium in China can be categorized into high, medium, and low suitable habitats, covering an area of 189.44×104 km2, 162.05×104 km2, and 252.87×104 km2, respectively. These areas account for 19.73%, 16.88%, and 26.34% of China’s total land area. The primary environmental factors influencing the distribution of X. sorbifolium, ranked from most to least influential, include annual precipitation, soil pH, average temperature of the coldest quarter, elevation, isothermality, precipitation coefficient of variation, soil organic carbon content, clay mass fraction, percentage of gravel volume, sand content, soil capacity, and maximum temperature of the warmest month. The results of this study provide a scientific basis for the selection and production planning layout of X. sorbifolium ecological cultivation bases.

At present, the planning of mobile routes plays a prominent role in ensuring the transportation of outdoor vehicles. With the advancement of technology and market drive, users have put forward higher requirements for the timeliness and accuracy of route planning. In this paper, a method of maneuvering route planning based on geographic and geological elements is studied on the basis of full consideration of maneuvering safety and various influencing factors in order to enhance the mobility protection ability in complex and harsh environments, and reduce unnecessary time and manpower consumption. First of all, comprehensively analyze all kinds of geographical and geological elements, assign scores and evaluate the key elements affecting the mobile route planning, and determine the weight of single element; Then, the data structure model is established to quantify the geographical and geological elements involved, and complete a multifactor overlay analysis; Finally, based on the A* algorithm to complete the construction of the motorized route planning model, to achieve the optimal route planning from the starting point to the end point. The experiments shows that: the motorized route planning method fully considers the influence of geological elements on motorized access on the basis of the traditional consideration of geographic elements only, and the model is more scientific and reasonable. In the case of known vehicle parameters, the scientific planning of maneuvering routes can be quickly completed, which has high market application value.

Image noise will inevitably be produced during the acquisition and transmission of digital images due to environmental interference or device component failure. Salt-and-Pepper Noise (SPN) is a typical sort of image noise. The noisy pixels, which can be described as randomly distributed white and black pixels on the image when SPN corrupts it, have the maximum and minimum values. SPN not only reduces the image quality but also drastically impairs the ability to retrieve image edge details. To enhance the image’s quality, it is crucial to use a top-notch image denoising method. The Euclidean-Distance-based Switching Adaptive Mean Filter (EDSAMeanF) is a new algorithm introduced in this paper. This technique splits the pixels into two groups: noisy pixels and regular pixels. The next step is to process each noisy pixel individually in order. Its templates are built using the Euclidean distance and are adaptive based on the noise density. This method is distinguished by its simple structure, quick execution, and robustness. Finally, we test with 10 common images and evaluated EDSAMeanF against 9 cutting-edge filters in terms of SSIM, MS-SSIM and PSNR. The result reveals that when the noise density is less then 50%, the denoising effect of EDSAMeanF is the state-of-the-art.

Under global warming, the increase in extreme rainfall events has led to a significant rise in the frequency of urban waterlogging disasters, and major cities are facing severe risks of waterlogging. In this paper, the revised “annual maximum method” was utilized to develop the rainfall intensity formula suitable for Xicen Science and Technology Innovation Center (XICEN). Based on satellite remote sensing data and SCS-CN model, we evaluated the urban waterlogging risks under five extreme rainfall scenarios with return periods of 5a, 10a, 20a, 50a, and 100a through the inversion method of critical rainfall. The results indicate that when the hourly rainfall exceeds 50mm, waterlogging may affect the entire study area. Furthermore, as return period extends, the risk of waterlogging also increases. Under the scenario of 5a return period, the waterlogging risk is mainly categorized as mid-low or low risk; however, under the scenario of 100a return period, the waterlogging risk of XICEN exceeds mid-high levels. These findings can serve as references for urban waterlogging disaster response and future construction planning.

In order to improve the population grid data of Gansu Earthquake Agency, this paper takes Liangzhou District as the research area. According to the principle of 'there are people in houses, no people in houses', a linear model is established by using the area of single houses and night light VNP46A2. Through comparison with the area method, it is found that the population distribution obtained by the linear model with night light is more consistent with the actual situation. Moreover, the accuracy is also higher than that of the population data obtained only by the calculation of the building area, which indicates the feasibility of the night light VNP46A2 data in the economically developed township scale in Gansu Province, and finds a simple and high portability method for the township level fine population spatial work in Gansu Province.

The calibration of transformers is a necessary aspect of maintaining the stability of the electrical power system. Nonstandard transformer calibrations are usually realized manually. However, manual calibrations are inefficient and inaccurate. To realize efficient and robust automated transformer calibration, we propose an edge feature-based template matching algorithm to detect the bolts of the voltage transformer. This algorithm combines the advantages of template matching and edge detection algorithms and is suitable for visual inspection in practical projects. This study provides a useful discussion for the reliable detection and identification of voltage transformer bolts. In the future, the algorithm parameters can be further optimized to adapt to a broader range of scenarios and application requirements.

Aiming at the problem of limited destriping performance caused by ignoring the non-local similarity of strips in existing methods, a hyperspectral image (HSI) destriping method based on non-local low rank tensor decomposition and total variation is proposed. This method considers the nonlocal similarity of the strips, and uses the tensor nonlocal low rank decomposition to approximate the nonlocal strips. Combining the direction and structural sparse features of the strips, they are effectively separated from the degraded image. Meanwhile, the local and nonlocal similarity of HSIs are considered together to reduce spectral distortion. Simulation results show that the proposed method accurately estimates and separates the strips, recovers the image information affected by the strips, and overcomes the problem of strip residue by considering the non-local similarity of strips and images.

In the field of ultra-high-resolution image segmentation, challenges persist regarding high resource consumption. Prior research, exemplified by ISDNet, introduced an innovative dual-branch network and a relation-aware feature fusion module. This module was designed to integrate shallow and deep features in a novel manner, resulting in improved inference speed while maintaining accurate segmentation. Our study aims to address the limitations of the original research and introduce enhancements. In this paper, we propose RSDNet, a Dual-Branch Feature Enhancement Framework for Super-Resolution Segmentation. Specifically, in our approach, the feature fusion module of ISDNet was replaced with a feature enhancement module. This module employs a feature addition method to combine deep and shallow features. Experimental results on the Inria Aerial dataset demonstrate a substantial improvement in the trade-off between memory usage and accuracy compared to the original method.

Medical image segmentation is a crucial task within the realm of medical image processing. Nevertheless, the intrinsic characteristics of medical images and the limited availability of data constrain the model's generalization capacity. Addressing this challenge requires an infusion of more data and the implementation of effective segmentation techniques to enhance model performance. In response to this need, we propose a deep attention enhanced network for medical image segmentation. This innovative approach boosts the segmentation model's efficacy through techniques such as data augmentation, a deep attention enhanced decoder, and a dual convolutional segmentation head. Validation across multiple datasets substantiates the method's effectiveness and its robust generalization capabilities.

In the field of medical imaging, C-arm systems play a pivotal role in surgeries, especially interventional surgeries. However, the current C-arm imaging system cannot adapt to the application scenarios due to algorithm limitations. Therefore, in order to improve the quality of diagnostic and therapeutic interventions, it is necessary to improve the quality of C-arm imaging and to fulfill the various requirements of different scenarios by improving the image processing techniques associated with C-arm imaging. This article explores how to improve C-arm image processing using convolutional neural network (CNN)-related algorithms. CNNs are known for their ability to extract complex features from images, which can be used for contrast enhancement, noise reduction, artifact removal, and resolution enhancement. Through a comprehensive review of state-of-the-art applications and technologies, this study proposes an innovative method of applying CNN for image processing in C-arm systems, improving image quality and thus providing potential advancements in clinical diagnosis and surgery. In terms of real-time image enhancement, integrating CNN into the C-arm imaging pipeline has demonstrated promising results, providing a reference for medical imaging learning and research. This article discusses the methods, experimental results, and implications of incorporating CNN into C-arm imaging, emphasizing that this technology can greatly improve image quality and indicating the potential of artificial intelligence technology in C-arm imaging applications.

In order to improve the effect and quality of video surveillance, we hope to use cutting-edge technology to realize video event recognition and other functions, and enhance the ability to respond to emergencies. Firstly, this paper briefly introduces the network framework, data enhancement processing method and loss function of YOLO series algorithm and YOLOv5 model. Further, it briefly explains the improvement of the improved YOLOv5 in this paper compared with the previous YOLOv5 version, including the enhancement strategy of the data set, the equalization processing strategy of the data set, the processing of the data characteristics, the improvement of the backbone network, and the tuning of the loss function. Finally, the number of classified samples of the training data set is introduced and the experimental results of the trained YOLOv5 are analyzed and visualized.

The defect inversion technology based on guided waves can detect defects in plates and pipes, including the location and shape information. An ultrasonic guided wave defect inversion imaging method based on fluctuation fields is proposed in this paper. Firstly, the principle of beamforming (BF) imaging is derived, which is also implemented by using simulation software for defects inversion. Then, through the mapping relationship between BF and Fourier diffraction theorem (DT) in the frequency domain, the results of BF are transformed into DT images, and a clearer inversion image is obtained. The inversion results demonstrate that this method can accurately reconstruct the location, size and shape of the thinning defects in the aluminum plate more with high resolutions.

In recent years, digital simulation has gained widespread application due to its ability to simulate target images in complex environments (interference) without expending significant human resources and financial investments. The simulation results can serve as the experimental foundation for algorithm design and system evaluation. Digital simulation of aircraft targets' infrared images involves the use of computer graphics and infrared radiation theory to visualize the physical infrared radiation. This paper presents a simulation methodology that, based on the creation of a typical aerospace background and an aircraft model, takes into consideration background radiation and atmospheric transmission during flight. This approach enables the generation of more realistic infrared scenes. It employs Fluent to solve the aircraft's temperature distribution, utilizes a program based on the Reverse Monte Carlo method to calculate the radiative intensity of aircraft surface elements, simulates the principles of an infrared detection system, and produces infrared images of the aircraft target from specific viewing angles.

This paper focuses on super-resolution of natural scene text images, serving as a crucial preprocessing step for image recognition, with the aim of extracting text information from low-resolution images. In this preprocessing step, the primary challenges stem from the impact of factors such as lighting, occlusion, and distortion on text features in natural scenes, as well as the tendency of small pixel values in low-resolution images to be overlooked by large models. Previous works have primarily focused on the capability to extract features. However, the key to addressing this problem lies in restoring the different distributions between high-resolution and low-resolution images. As widely acknowledged, diffusion models excel in learning to generate distributions from datasets. Therefore, we introduce a conditional diffusion model to learn feature distributions among diverse images and apply histogram matching to address the issue of n patches. Our experimental results confirm the effectiveness of our approach.

With the update iteration of deep learning technology, the necessity of image similarity calculation in image retrieval, target detection and tracking has become more and more prominent, and it has become an important force to promote the development of computer vision and artificial intelligence. In the face of increasingly complex and diverse application scenarios, the traditional image similarity detection technology has obvious shortcomings in calculation accuracy, and the image feature extraction needs manual assistance in the calculation process, so that the overall robustness of the technology is poor and cannot meet the practical application requirements. In view of this, this paper integrates the powerful image feature extraction ability of convolutional neural network, and constructs an image similarity calculation model based on VGG16- Siamese neural network, aiming at improving the accuracy of image similarity calculation and providing data support for the next technical application. Practice has proved that the VGG16 part of the model is used as the backbone feature extraction network to obtain the feature vector of the image. The Siamese neural network can use the feature vectors of two images to construct the distance similarity function at the same time, so as to calculate the similarity of the two images. The final experimental results show that the image similarity calculation model of VGG16- Siamese neural network has a precision of 90.84% and a recall rate of 91.88%, which meets the design expectation and has certain application and promotion value.

In addressing the issue of unstable boundary adherence in traditional superpixel segmentation algorithms, this paper proposes a novel region-adaptive superpixel segmentation algorithm based on density clustering for computer vision research. Firstly, during the initialization of superpixels, a region-adaptive initialization method is employed based on the complexity of image content. This method takes into consideration the diversity of image content, thereby reducing redundancy in the initialization to a certain extent. Secondly, within the initialized superpixel regions, two pixels properties, namely distance and membership relationship, are computed. Pixels with higher local density are chosen as seed points, and their membership relationships are marked. Finally, seed points closest to the pixels are selected in the membership relationships, and labels are assigned to the pixels, completing the superpixel segmentation of the image. Experimental results, compared with six advanced algorithms, demonstrate that the proposed algorithm exhibits superior performance in terms of boundary recall rate, under-segmentation error rate, and accuracy while maintaining a higher level of edge adherence.

While recent advancements in image inpainting techniques have shown significant improvements, it often faces challenges in producing realistic image structures, particularly when it comes to filling large holes within intricate images. Moreover, considering computational efficiency, it is common practice to train the network using low-resolution images. To address the restoration challenge of high-resolution images with significant missing regions, a new method called Multi-Conv-Transformer is proposed in this paper. We integrate the advantages of Transformer and CNNs to enable the model to exhibit efficient performance in training high-resolution images. A customized transformer block specifically optimized for inpainting purposes is introduced in our work. Within this block, the proposed multi-head self-attention module collects non-local information only from valid tokens identified by a dynamic mask., thus prioritizing the partial regions of the image. The experiment results demonstrate that our model performs well in the restoration of various scenarios.

In image retrieval, two main indicators are focused: accuracy and efficiency. Focusing on improving the indicators, this article proposes an optimized method for image retrieval based on CNN features based on Fourier transform and entropy. Using Fourier transform to describe images more accurately to improve accuracy, Using entropy to binary describe images to improve efficiency. We evaluated the effectiveness of our method using Vgg16, Resnet18, and ShuffleNetV2 networks on the UKbench, Holidays, and Wang datasets. On UKbench, the average improvement in retrieval accuracy is 0.0254, and the average improvement in retrieval efficiency is 22.12%. On Holidays, the average retrieval accuracy improved by 2.65% and the average retrieval efficiency improved by 18.63%. On Wang (top 20), the average retrieval accuracy improved by 0.23%, and the average retrieval efficiency improved by 37.58%. The experimental results show that the proposed method can effectively improve accuracy and efficiency.

With the continuous development of digital photography technology and quality inspection, the demand for image stitching in practical applications is increasing. Traditional image stitching algorithms employ a variety of hand-designed methods for feature extraction, matching, and optimization. However, these traditional feature-based image stitching techniques heavily rely on feature extraction and may not perform well in scenarios with limited features. Current image stitching solutions based on supervised deep learning lack relevant data sets, and labeling data is relatively cumbersome, making supervised deep learning methods unreliable. At the same time, the rise of unsupervised deep learning algorithms provides new ideas for image stitching. We use unsupervised homography estimation to provide information about the geometric relationship between images, Stitching-Domain Transformer Layer to align feature maps, warp and generate masks, it helps to enhance the reality and continuity of splicing. We simultaneously utilize a pre-trained deep learning model (VGG) for feature extraction. We adjust the smoothness loss term to ensure smoother transitions within the stitching areas. Throughout the training process, we continuously optimize the number of convolutional layers, channels, and network depth to achieve optimal results. The superiority of the unsupervised learning algorithm compared to other classic algorithms was verified through experiments. Finally, we discussed the challenges and future applications of unsupervised deep learning in image stitching.

Low-light image enhancement is for the sake of improving the quality of visual perception of captured data on the conditions under low-light to obtain more information, and has gradually become a research hotspot in the field of image processing. It has broad application prospects in artificial intelligence related industries such as autonomous driving and security. This article reviews the current research status of low-light image enhancement algorithms, provides a detailed analysis of the circumstances of low-light image enhancement algorithms, and analyzes the current low-light image enhancement algorithms. Based on the exploration of the above current situation and combined with practical applications, this article points out the limitations of the current algorithm and predicts its development trend.

Traditionally, the correction of panoramic images has been achieved through sundial projection or cubic projection. However, these methods have not adequately addressed the issues of overlap and fixed pose relationships in the corrected images. To address this, we propose an enhanced algorithm for correcting panoramic images using cubic projection. The transformation process involves mapping the spherical panoramic image onto a cube that is tangent to the panoramic sphere, with the cube's side length exceeding the diameter of the sphere. For the re-projection correction experiments, we selected five sets of panoramic images and compared feature extraction and matching using the SuperPoint, SuperGlue, and R2D2 algorithms. The experimental results demonstrate that our proposed method outperforms existing methods in terms of image re-projection, showing significant enhancements in image overlap, accuracy, and the number of feature matches.

With the rapid advancement of synthetic aperture radar (SAR) sensors, it has become more important to extract change information between high-resolution SAR images. Considering the efficacy and robustness of segmentation-based strategy in solving fine-resolution images, as well as the excellent performance of deep learning in feature extraction, therefore, in this paper, we propose an unsupervised region-level graph convolutional network (URGCN) for SAR change detection. Firstly, we propose a multi-temporal joint segmentation method to generate multi-level regions. This method can simultaneously segment bi-temporal SAR images and automatically align bi-temporal regions during segmentation process. We then perform pre-classification on difference images generated by neighborhood ratio detector. The high confidence training samples are selected through hierarchical clustering. Thirdly, we encode the segmented regions are as a graph structure and construct GCN model. The graph can capture the spatial and temporal structure information. Finally, the GCN model parameters are optimized by labeled samples and inferring unlabeled regions. Experimental results on two VHR SAR change detection datasets demonstrated that the proposed method can extract complete change information with high accuracy compared to alternative approaches.

Keypoint detection is a crucial step in feature-based image registration methods. Traditional approaches typically rely on corner or blob matching. In this paper, we address the registration problem of optical and synthetic aperture radar (SAR) images by proposing the use of phase consistency maps combined with maximum moment maps for image enhancement. We then perform joint detection of multi-scale corners and blobs, followed by subsequent processing, in order to improve the repeatability and spatial distribution of the detected key points. By combining these two types of keypoints in the proposed method, we can benefit from the complementary strengths of corner and blob detection. This approach allows us to capture a wider range of salient image features and improve the robustness and accuracy of keypoint detection.

The transmission line conductors are an essential component of power system operation. Due to prolonged exposure to the external environment, these conductors are vulnerable to damage caused by various factors, leading to power system failures that have significant impacts on production and daily life. To monitor the condition of transmission conductors, assess the extent of damage, and ensure stable and safe grid operation, we propose an intelligent wire damage identification technology based on transmission wire damage images. Initially, we apply gray variance normalization and median filtering techniques to enhance the quality of the transmission wire image. Subsequently, local threshold segmentation is performed on the enhanced image to extract the wire area. Leveraging the characteristic that damaged areas exhibit higher gray values compared to other regions, we employ a gray value vertical and horizontal projection algorithm along with subsequent filtering processes for identifying surface damages on wires. This approach effectively transforms wire damage recognition into a data processing problem. Finally, through a series of experiments conducted in power maintenance operations, we validate its feasibility.

The road network is an important reference standard in infrastructure planning. Whether the road network is complete or not better reflects the economic and humanistic conditions of urban construction. In this study, the current severe challenges in drone remote sensing road network extraction are addressed, such as low extraction efficiency and weak generalization performance. An end-to-end unsupervised road network extraction framework based on drone images is proposed. In comparison with the results of the best-supervised learning methods and unsupervised learning methods, the overall performance shows the best performance. Among them, the IOU and F1 scores reached 0.95 and 0.96, respectively. The method proposed in this study provides a new reference for feature extraction from drone images and can help with the real-time update of the road network.

To improve the accuracy and efficiency of detecting occluded areas in large tilt images, this paper improves the Occlusion Detection Algorithm Based on Elevation Constraints on the basis of Angle Based Occlusion Detection Algorithm, and proposes an Occlusion Detection Algorithm Based on Elevation and Angle Constraints. This algorithm uses the spiral scanning method to judge the visibility of ground points layer by layer from the underground point, and fully utilizes the elevation, elevation angle with projection center, and visibility information of the inner layer point to judge the current point, greatly reducing the number of points that need to be judged. Compared with existing classical algorithms, it has the advantages of high accuracy and fast visibility analysis speed.

To address the issue of We-maps users experiencing difficulty in quickly and accurately retrieving maps, the paper proposes utilizing perceptual hash algorithms to retrieve the maps. First, the map image is extracted from the We-map database and processed in batches to obtain invariant features. Then, various perceptual hash algorithms are used to establish spatial mapping based on map content features, resulting in a fixed-length hash code. Finally, the Hamming distance is used to calculate the similarity of hash features corresponding to the map, and a comparison is made to retrieve the map from the We-map database that fulfills the user's requirements based on the comparison outcome. The retrieved map, based on the comparison result, will satisfy the user's needs. Using this method enhances image retrieval efficiency, enabling fast retrieval of images on We-map and comprehensive micro-map image retrieval. The retrieval-based experimental results demonstrate that by utilizing this approach, users can promptly retrieve We-maps with comparable image characteristics, thereby presenting an innovative content retrieval method for We-maps.

Panoramic annular image is an optical image that projects surrounding objects onto a circular area with the lens center as the focal point. It is widely used in petroleum pipeline monitoring system, but it is difficult to achieve high-quality images due to complex natural conditions in the surroundings, which affect the accuracy of pipeline monitoring system. To address this issue, we designed an optical image system which can obtain self-collected datasets. A dual-path network super-resolution algorithm is proposed, which utilizes an enhanced deep recursion network to extract detailed feature information from images, an attention mechanism network is utilized to extract important feature information from image in another path. Then the information of two paths is fused together to reconstruct high-resolution image. Experimental analysis and comparison were conducted on self-collected datasets using the algorithm, and the results demonstrated that our approach significantly improves image details and contributes to enhancing the accuracy of pipeline monitoring system.

Medical image classification plays a crucial role in clinical treatment and medical education. However, traditional classification methods have reached performance limits and require substantial time and effort for feature extraction and selection. In this study, we introduce a novel deep learning model that combines the feature extraction part of DenseNet with the dynamic routing mechanism of capsule networks. DenseNet, known for its dense connections, aids in gradient propagation and feature reuse. Capsule networks introduce capsule layers with sensitivity to spatial relationships. By merging these two structures, our aim is to enhance the model's ability to model object hierarchy and pose variations. Our approach initially incorporates the dynamic routing mechanism of capsule networks on top of DenseNet, achieving an effective fusion of the two structures. Experimental results on the Chest X-Ray14 dataset demonstrate a significant improvement in accuracy and robustness compared to traditional DenseNet and capsule networks. To address the issue of limited annotation in medical image data, we introduce a semi-supervised learning approach and improve training balance through anti-curriculum pseudo-labeling. Furthermore, to obtain more accurate pseudo-labels, we adopt a mixed model for pseudo-label generation. This method not only effectively balances the pseudo-label predictions generated by two models but also mitigates the challenges associated with threshold-based confidence intervals in the context of multi-class, multi-label scenarios. In summary, the combination of deep convolutional neural networks and capsule networks, along with semi-supervised learning and anti-curriculum pseudo-labeling, holds great promise for medical image classification. It has significant application value in clinical practice and presents a wide-ranging prospect for the future.

The scattering effect of random media such as clouds, turbid water, and biological tissue severely restricts the imaging and ranging capabilities of optical detection and ranging systems through media. The method of using precise simulation of the light transmission process to invert the target image has gradually become the mainstream method for imaging through scattering media. In the diffusion equation method commonly used for simulating light transmission in random media, it is necessary to strictly determine the boundary of the random media, which is difficult to achieve in practical applications. We propose a boundless point spread function for simulating the propagation of light in random media, which equates the diffusion of light in a finite-sized random media to that in an infinitely large random media, without obtain the boundary of the random media. Our proposed imaging method may have potential applications in fields such as medicine, rescue and relief, exploration, and autonomous driving, where obtaining information about the boundary of random media is typically difficult.

The optimization of image recognition technology provides a platform and possibility for the dissemination and protection of most of the information documents, especially font and number recognition, which is an important part of it. At present, the recognition accuracy of standard fonts and numerals has reached its maximum with continuous optimization, but the recognition of handwritten numerals still needs to be further improved. The aim of this paper is to test the correct recognition rate of handwritten numerals using a convolutional neural network model. Through the elaboration of the basic principles and structure of the neural convolutional network algorithm and the purpose of the article, the relevant test model was established. And the MINST dataset is used to train the model and test the recognition effect of handwritten numbers. Through experiments, the model used in this paper can correctly recognize most handwritten digits, thus satisfying the purpose of the experimental design, and providing new support for experiments on handwritten fonts.

In recent years, the damage of X-rays to the human body has received widespread attention. This paper analyzes and improves the image super-resolution reconstruction algorithm based on adversarial-generative networks that performs well in the field of image post-processing. This algorithm can largely weaken the The loss function in the convolutional network directly manages the defects of the result, thereby achieving the purpose of restoring image details more clearly. We have purposefully added structures such as filtering, feature extraction, and detail restoration to the generation network to achieve a state suitable for low-dose CT image restoration. Experimental results show that this kind of network has a more stable reconstruction effect and can obtain clearer texture and detail information.

With the rapid development of deep learning, the task of image super-resolution has made significant progress. However, as model depth increases, training becomes more difficult and fails to capture coarse-grained and fine-grained information simultaneously. To solve these issues, we propose Deep Hierarchical Multiscale Attention Networks (DHMA). First, we use a residual nested residual structure to improve the propagation of gradient information to address the challenge of training deep networks, which consists of multiple residual groups, each composed of multiple residual blocks. In addition, we propose a Hierarchical Multiscale Attention (HMA) module to capture both coarse-grained and fine-grained features in deep networks. This method imitates how humans observe things, first focusing on salient objects and then observing details. Specifically, we first segment the feature map horizontally into several parts and then perform Global Aware Attention (GAA) learning on each part. Where GAA can learn global structural information. Next, we use the Adaptive Feature Fusion (AFF) module to fuse the learned information of each part into the current layer's features. Finally, we stack the features of each layer to construct a hierarchical multiscale structure and obtain features of different scales. HMA is a lightweight plug-and-play module that can be applied to existing models. Extensive experiments demonstrate the effectiveness and outstanding performance of the DHMA.

To solve the problems of multiple incorrect matching points, low matching accuracy, and slow matching speed in visible light image stitching of UAV. This article proposes an improved whale algorithm feature matching method based on quaternion representation of color images. The classic ORB feature extraction method has been improved by introducing quaternions to represent color images, and using the improved whale algorithm to find feature points where quaternions represent color images. The direction between the centroid of the quaternion moment and the quaternion feature point is taken as a direction to solve the problem of lack of direction in quaternion key points. Using a feature point attachment degree method to purify the obtained feature points for feature matching in UAV color images. The experimental results show that the proposed feature matching method for UAV color images based on quaternion improved whale algorithm is the best among the four experimental methods.

Deep learning image compression, using neural networks, improves compression over traditional methods like JPEG. These methods enhance visual quality at lower bit rates by learning better image representations. However, they struggle with capturing broad context compared to local features. To address this, we propose enhancements: a new convolutional module with stacked layers and advanced operations, and a spatial attention block ("Shuffle attention") for better feature extraction. These boost performance. Our method is faster and requires fewer parameters than state-of-the-art techniques on Kodak and CLIC datasets. Despite slightly lower rate-distortion performance, our Composite Conv module and spatial attention block effectively extract global features and reduce encoding time. In conclusion, our work advances deep learning image compression by mitigating convolutional network limitations, enhancing compression efficiency while preserving quality.

The Low Light Image Enhancement (LLIE) task aims to restore images with poor lighting conditions and visual effects to images with good lighting conditions and visual effects. However, the enhancement results output by existing methods always include various image degradation such as overexposure, artifacts, and color shift. To alleviate these issues, a low light image enhancement method based on Retinex theory and Transformer is proposed. Designed a dual branch network architecture with Transformer blocks. One branch is used to enhance local details of the image, while the other branch is used to adjust the color and brightness of the enhancement results. The dataset used in the experiment was the LOL dataset, which was 23.42 and 0.81 in the PSNR and SSIM indicators, respectively. Compared with the suboptimal algorithm, it improved by 1.06 and 0.016, respectively. The experimental results show that this method has achieved excellent performance in low light image enhancement tasks and has certain application value.

Deep learning has made significant advancements in the field of computer vision in recent years. However, research in target detection technology has primarily focused on improving the data reuse rate, model generalization capability, and detection accuracy. In this study, we propose a data improvement technique called CutMix-Plus, which utilizes the YOLOv5n baseline network. We validate this technique using publicly available datasets. The mAP@0.5 of this algorithm is 0.892, and the mAP@0.5-0.95 is 0.548. These values represent improvements of 7.6% and 4.8% for mAP@0.5, and 5.6% and 4.6% for mAP@0.5-0.95, respectively, when compared to the CutMix and Mosaic algorithms. The experimental findings demonstrate that the approach can significantly improve the model's ability to generalize and improve its accuracy. Additionally, it can meet the requirements for detecting helmet wearing in construction scenarios.

Segmentation on medical image is a necessary prerequisite for disease diagnosis and treatment planning. In various medical image segmentation tasks, U-Net based on convolutional neural network (CNN) has achieved tremendous success due to its ability to learn image details and deep high-dimensional features. However, the inherent limitations of convolution operations limit their performance in modeling explicit long-term relationships. Especially in the brain MRI glioma segmentation task, traditional convolutional neural networks show weakness in learning global semantic information due to the extreme intrinsic heterogeneity of tumors in terms of appearance, shape, and histology. Therefore, this research proposes a 2D U-Net combined with transformer for brain tumor segmentation, and also designs a fusion module to better fuse the high-resolution detail features from the encoder with the high-level semantic features in the decoding process. The respective segmentation models are trained on the three orthogonal planes, and a tumor core-assisted network is separately trained. We validate proposed method on the BraTS20 brain glioma MRI dataset, and the experimental results demonstrate that our method outperforms state-of-the-art 2D methods. Compared with 3D methods, our model can accomplish accurate tumor segmentation while saving computational resources.

With the rapid development of water resources, the health problems of underwater facilities have attracted more and more attention. Aiming at the surface health problems of underwater facilities and avoiding the requirements of optical methods on light and water quality, research on underwater acoustic images was carried out. The fuzzy theory and membership function were used to filter pixels, and the features in acoustic images were obtained and the typical features in traditional acoustic images were compared and analyzed. Compared with traditional methods, fuzzy theory can extract the typical features of images better.

In response to the current contradiction of rapid growth in lung CT image data and insufficient human diagnostic capacity, we propose a medical image discrimination method based on an improved 3D-CNN model. This method first segments the problematic regions in the images. Then, it utilizes a category-weighted improved loss function and increases the convolutional layers to construct a 3D-CNN model. Finally, the model is fine-tuned to predict the probability of malignancy for tumors. Additionally, we leverage repeated sampling and data augmentation to address data imbalance issues, thereby enhancing the effectiveness of the model. The experimental results indicate that this discrimination method, compared to the original 3D-CNN model, has shown a significant improvement. The F1-score has increased by 27.3%, Precision by 34.5%, and Recall by 13.8%. It can more effectively detect pulmonary nodules, thereby reducing the areas that doctors need to assess and enhancing overall work efficiency.

The microdroplet jetting speed of dispensing valve is very fast, generally tens of milliseconds, microdroplet diameter size is generally only a few tens of microns. In order to study the consistency of the dispensing valve jet, this paper analyzes the forming process of microdroplet and microdroplet volume calculation method, builds an observation system, realizes the non-contact measurement of microdroplet volume based on image processing, and confirms the feasibility of the method by a high-precision electronic balance.

In this paper, a post-processing method to automatically detect and characterize ionospheric plasma irregularities in low latitudes has been proposed. When monitoring the ionosphere using the total electron content (TEC) calculated by GNSS beacons, plasma bubbles appear as TEC depletions and plasma masses appear as TEC enhancements. Therefore, plasma irregularities can be detected and characterized by identifying depletions and enhancements of TEC. The method is based on an analysis of the rate of change of relative total electron content (RTEC), followed by a high-order polynomial fitting strategy to extract short-time series in which TEC depletion (enhancement) does not occur. General characteristics of ionospheric TEC depletions or enhancements, such as start and end times and magnitude, can be automatically provided by the method. The ionospheric irregularity seeker based on GNU Octave programming language has been developed based on this method, and the analysis of beacons data of different stations in low latitudes with main sampling rates (e.g., 1s, 15s, 30S) of Beidou, GPS, and GLONASS has been realized. The reliability and practicability of the method are verified by using the GNSS measurement data provided by the HKCL and HKQT stations in the equatorial anomaly region of China from August 2013 to December 2015. It was also used to analyze the measurement data from HKCL stations and HKQT stations in October 2023. Compared with previous studies done with different techniques and measurements, these results showed clear agreement.

We identified two challenges in the current creation of depth-annotated dataset for autonomous driving or 3D reconstruction: 1) The viewing angle difference between cameras and radar, causes projecting background points onto foreground objects in the Lidar-to-image projection process, leading to inaccurate depth information acquisition. 2) The substantial manpower required to eliminate occluded points during the depth dataset creation process. In addressing these challenges, we analyzed the principles behind occlusion and proposed an automated filtering method for point cloud projection images. Our method involves segmenting images into various-sized sliding windows based on different scenes to filter out occluded points. Through this approach, we achieve effective occlusion removal in diverse scenarios, encompassing structured objects like cars, buildings, pedestrians, as well as unstructured objects like trees. Moreover, the proposed algorithm automates the process, thereby significantly reducing both labor and time costs. Additionally, we proposed different schemes for creating sparse and dense depth maps based on the density of the camera depth image. We have open-sourced the code, and the repository can be found at: squirreljj/An-efficient-framework-for-creating-depthannotated-image-datasets-with-camera-and-Lidar-fusion- (github.com)

This article mainly studies the principle of multilayer perceptron (MLP), including the forward propagation and backward propagation processes, and how to use Tensorflow to implement multilayer perceptron for image recognition. First, we outlined the principle of multilayer perceptron and explained the forward propagation and backward propagation processes in detail. Then, we used Tensorflow to build a multilayer perceptron model and conducted training and testing. Finally, we applied the model to image recognition tasks and achieved good results.

In the process of acceptance measurement based on drone point cloud data, the mining cars in the open-pit mining area will generate a large number of pseudo terrain points in the 3D point cloud modeling process. If they are not extracted and removed, the accuracy of acceptance measurement will be seriously reduced. Therefore, how to accurately identify and extract mining car points from 3D point cloud data is the main technical issue for high-precision acceptance of mining sites based on unmanned aerial car oblique photogrammetry technology. Therefore, this article proposes a machine learning algorithm based method for extracting open-pit mining car point sets, which can effectively solve the problem of reduced acceptance accuracy caused by the existence of mining car point sets, thereby improving the accuracy of acceptance measurement. This provides important technical support for open-pit mining site acceptance measurement based on drone photogrammetry technology.

This paper introduces a deep learning method tailored for recognizing and classifying vegetables, targeting the challenges in identifying vegetables within the vegetable flow process. A limited dataset including five common vegetable types was assembled, divided into training and testing subsets. Data augmentation techniques were employed to balance the sample quantities. Based on the TensorFlow deep learning framework, a Convolutional Neural Network (CNN) was constructed comprising four convolutional layers and two fully connected layers. Techniques such as L2 regularization and Dropout were employed to mitigate overfitting. Additionally, a confusion matrix was utilized to visualize the predicted quantities and accuracy for each class of images. Unlike complex architectures handling extensive datasets, this model offers simplicity, enabling easy deployment on terminals. Its low power consumption and minimal hardware demands suit the needs of routine vegetable circulation processes. Experimental results reveal an enhanced recognition accuracy with larger dataset sizes. Adam as the optimization algorithm notably outperforms SGD and momentum in the experiments. Moreover, employing a greater number of convolutional kernels enhances recognition effectiveness. On a dataset comprising 175,536 training images and 43,884 testing images, the recognition accuracy consistently stabilizes above 80%. The peak recognition rate achieved reaches 84.13%.

In order to accurately identify the crossing behavior of power operators in the power grid dispatch operation area and maintain the safety of power grid operation, a technology for accurately identifying the crossing behavior of power operators in the power grid dispatch operation area is proposed. Collect monitoring video images of the power grid dispatch operation area, and segment the monitoring video images into target foreground and background based on information entropy; Extract time-domain and frequency-domain characteristics of cross behavior of power operators; Recognition of cross regional behavior in power grid dispatch operations based on local gradient feature histograms. The experimental results demonstrate that this method can improve recognition accuracy, recall, and efficiency, and contribute to maintaining the safe operation of the power grid.

In recent years, the development of object detection technology has become increasingly mature, and its applications in fields such as personnel detection, pedestrian detection, vehicle detection, and remote sensing imaging are becoming more and more widespread. With the development of video surveillance equipment, combining object detection with video surveillance equipment to achieve intelligent analysis applications based on classroom surveillance videos has important practical significance.

This paper proposes a comprehensive method for detecting maritime obstacles based on marine radar PPI images. This method not only utilizes radar imagery for the rapid identification and tracking of obstacles but also integrates ship GPS and bow direction information to measure the distance, speed, and direction of obstacles. Testing on actual ships has shown that this detection method, which fuses radar imagery with GPS information, can help seafarers observe a more comprehensive picture of the surface situation, thereby enhancing collision avoidance capabilities to a certain extent.

Crowd recognition by the deep learning is important to the management of campus or classroom, which include the technologies of object detection and face detection. In the crowd recognition case, the detection of anomaly samples plays key role like the person with no face or the opposite case. This study focuses on the cooperation of object detection model YOLOv8 and the advanced facial detection model SCNet, which is to find the anomaly samples. We try to provide a valuable insight into the research of object detection and facial recognition tasks in the field of computer vision, which is to provide useful guidance for further promoting the improvement of crowd recognition.

The existing YOLOv5s object detection model has issues of missing small targets and lower detection accuracy in complex scenes. Therefore, YOLOv5s needs to be optimized based on these problems. Firstly, the coordinate attention mechanism and the structure based on receptive field Block are introduced to improve the accuracy of the model. Secondly, the coordinate attention mechanism and the omni-dimensional dynamic convolutional structure are introduced to improve the situation of missing small targets. Selected part of the data in PASCAL VOC2007 for experiments, and the final results show that the model in this paper improves the recall by 1.7%, the average precision mAP50 by 2%, and the average precision mAP50-95 by 3% compared to the yolov5s model. The model in this paper improves the detection accuracy to a certain extent and improves the leakage detection of important targets and is tested in industrial scenarios with good results.

Through the revelation and internal logic relationship of the DIKW model, the logic of data, information, knowledge, and wisdom in small target detection research initially built a new point of view, and did some basic tasks for small target detection and research, so as to use it to use it to do some basic tasks for small target detection and research. Small target detection research model under DIKW. The DIKW Information System was introduced into the field of small target testing. First, the DIKW model was analyzed. The framework of data information-information knowledge-intelligent framework was studied, and small target detection methods of one-stage and two-stage stages were studied, such as CNN-based methods, transformer methods and mixed methods. This article analyzes the basic network framework of small target detection, and builds data, information, knowledge and wisdom in small targets detection research.

Patients with visual impairments often face challenges in identifying and navigating environmental obstacles in their daily lives, therefore reliable assistive technologies are needed to improve their quality of life and independence. This study proposes a visual obstacle assistive device guidance scheme that comprehensively utilizes sensor technology and object detection algorithms, aiming to provide comprehensive solutions for patients. We have adopted advanced sensor technology, including cameras, LiDAR and other sensors, to monitor the surrounding environment in real-time. These sensors can detect the distance, shape, and position information of obstacles and transmit this data to the navigation system. We use the YOLOv5 object detection model, which can accurately identify and track various obstacles, including pedestrians, vehicles, etc., under real-time camera input. The output of the target detection model is fused with sensor data to form a comprehensive environmental perception system. The experimental results indicate that the auxiliary device guidance scheme exhibits high accuracy and real-time performance under different environmental conditions, and has a positive impact on improving patients' quality of life and autonomy.

To address the issues of robustness and localization accuracy in traditional feature-based simultaneous localization and mapping (SLAM) systems in dynamic environments, we propose a real-time visual SLAM optimization method tailored for dynamic indoor scenarios. This paper builds on the ORB-SLAM2 algorithm framework and introduces a dynamic feature point pruning strategy. The server uses the YOLOv8 object detection algorithm to detect objects in the surrounding images and extracts object information such as bounding boxes and categories. The SLAM client continuously analyzes these results and removes feature points belonging to dynamic objects based on the detected prior information. It also eliminates potential dynamic feature points through epipolar geometry constraints. In cases where a small subset of feature points still satisfy the epipolar constraints but exhibit dynamic behavior, we compute the depth information difference between the current frame and the reference frame. If this difference exceeds a predefined threshold, the feature point is classified as dynamic. Finally, experiments conducted on the TUM dataset demonstrate that our algorithm effectively improves the localization accuracy and robustness of visual SLAM systems in highly dynamic datasets compared to the original ORB-SLAM2 algorithm, resulting in a remarkable 98.01% improvement in the absolute RMSE of the trajectory error.

Electronic medical records are important information for medical intelligence, although there have been many related studies. However, based on the characteristics of the electronic medical record itself, for example, there is no clear boundary of participle and other problems, which causes difficulties to the research. Chinese, as a kind of hieroglyphic script, contains rich information in itself. There are various methods to use the splitting of Chinese characters as an enhanced data input model to improve the overall recognition effect; however, there are various forms of splitting of glyphs, and there are no papers to compare and contrast this information enhancement method. In this paper, firstly, the common ways of Chinese NER are introduced, and then they are analyzed through the technical point of view, followed by illustrating the effects of different splitting methods on the NER task, and comparing this data enhancement method through experiments.

UAVs are used in various scientific and applied research due to their flexibility, compactness, intelligence and autonomy. Target detection problem for UAVs has gained many results in academia in recent years. In the target detection of UAVs, selecting the appropriate deep learning framework and targeting the functional refinement is one of the more important steps. In this paper, based on the principle of FGFA algorithm, we improve the FGFA (Fine-Grained Feature Aggregation) algorithm to accomplish the utilization of temporal information in the video of UAV target detection dataset. The model is benchmarked and compared with other target detection methods using publicly available UAV datasets. The experimental results show that the method in this paper has better performance.

With the continuous development of the automotive industry, the target detection and recognition of automotive parts have become crucial factors for automakers to enhance automation levels. In this paper, to ensure improved detection accuracy while meeting real-time detection requirements, we propose an enhanced YOLOv5 model for automotive parts recognition. This method optimizes candidate box parameters using the K-Means algorithm, addresses feature mismatch issues through variational convolution, and optimizes the loss function using GIOU instead of IOU. Experimental results show that the improved YOLOv5 algorithm achieves an average precision of 99.1%, representing a relative improvement of 1.72% over Fast R-CNN and 3.91% over the standard YOLOv5 algorithm. Furthermore, it demonstrates excellent real-time detection performance, meeting the demands of practical application scenarios.

UAV reconnaissance has many advantages, such as high reliability, strong flexibility and wide coverage. When using unmanned aerial vehicles to scout targets, it is often necessary to quickly scout many fixed targets. If it is obviously unreasonable to send one unmanned aerial vehicle for each target, it is also unreasonable to send one unmanned aerial vehicle to scout all targets. At the same time, we should also consider the actual needs, such as our own safety and the number of airport reconnaissance drones. Considering many factors, a fast reconnaissance model of multi-UAV fixed targets on the ground is established, and the minimum number of UAV sorties needed for reconnaissance of all targets is obtained by combining branch and bound method and modified combined optimization algorithm. The modified Dijkstra algorithm is used to determine the appropriate delivery point, which can meet the actual requirements of rapid reconnaissance in the battlefield, considering the factors such as air defense firepower range and transportation conditions. Finally, the feasibility of the model is verified by using simulation data, and the benefits and possible improvements of the model are analyzed.

Aiming at the problem of poor detection effect of MobileNet-SSD algorithm when the target is at the edge of the image or the target is small in human target detection, this paper proposes a human target detection method based on MobileNet- SSD combined with frame difference method. Firstly, the coarse position of the moving target is obtained based on the frame difference method; Secondly, the motion region of the target is calculated based on the rough position, and the region is intercepted in the detection image to achieve the acquisition of the adaptive region of interest (ROI), and then the ROI is sent into the MobileNet-SSD model to achieve human target detection; Finally, using the human target detection method based on MobileNet-SSD combined with frame difference method to carry out the human target detection experiment. The results show that the human target detection method based on MobileNet-SSD combined with frame difference method can effectively detect the human target in the image edge region without affecting the original detection speed.

Aiming at the complexity and large parameter size of you only look once version four (YOLO v4) object detection network, which cannot meet the requirements of lightweight deployment and real-time computation on mobile devices, a lightweight object detection network called Ghost-YOLONet is proposed. Firstly, the backbone network of YOLO v4 is replaced by the lightweight network GhostNet, and the decoupled fully-connected attention mechanism is integrated into the Ghost module to better capture global information. Then, the parallel structure of the spatial pyramid pooling module in YOLO v4 is changed to a serial structure to improve the model's execution efficiency. Comparative experimental results on the PASCAL VOC2007 and VOC2012 datasets show that compared with the YOLO v4 model, Ghost-YOLONet reduces the parameter size by 81.1%, the model volume by 81.5%, and achieves mAP@0.5 of 81.4%. Moreover, the FPS is improved, meeting the requirements of real-time detection tasks.

In response to the current challenges of poor real-time performance and low accuracy in facial expression recognition in classroom environments, this paper proposes a classroom facial recognition model based on YOLOv4. The original backbone feature extraction network is replaced with DenseNet121, and the SE attention mechanism is incorporated into the model to enhance feature extraction and representation capabilities. The EIoU is used for localization loss calculation to improve the localization accuracy of small targets. The improved model enables real-time prediction of multiple facial expressions. Validation on a self-built dataset demonstrates that the improved model achieves a 16.33% increase in mAP compared to other models, with an AP of 96.51% for commonly occurring facial expressions in the classroom. The model achieves a video inference speed of up to 25fps and maintains good recognition performance in online teaching scenarios, indicating a certain level of generalization.

Tracking targets accurately and robustly in visually complex environments poses a formidable challenge. To address this, capturing a resilient appearance representation is essential while augmenting the model's ability to generalize and cope with diverse challenges such as object deformation, variations in illumination, changes in scale, and motion blur. This paper presents a method for sturdy tracking in intricate scenarios, employing the efficient convolution operator (ECO) tracker. Our approach incorporates the 2 key concepts: a) extracting profound features via the Conformer network by increasing the number of underlying channels, and b) flexibly adjusting the fusion weight for shallow and profound features based on factors such as the peak-to-sidelobe ratio and the joint score of adjacent frame trajectory smoothness. This approach enhances the model's adaptability and generalization in intricate environments by capitalizing on the complementary aspects of deeper-layer and shallow-layer features. Experimental outcomes validate the algorithm's efficacy in tackling varied challenges related to target tracking in complex environments, ensuring robust tracking with consistently high accuracy.

Accurate target detection under heavy background noise and clutter remains the key challenge for radar applications. Traditional target detection methods based on statistical models have limited performance in complex environment. Although advanced deep learning methods such as Convolutional Neural Networks (CNN) have been introduced to suppress clutter and improve the target detection performance, the limited interpretability and flexibility hinders their further implication and extension. To address this challenge, A novel multi-dimensional feature target detection method based on deep forest (DF) is proposed in this paper. Firstly, a low threshold Constant False Alarm Rate (CFAR) detector is employed to preprocess the received radar signals. Afterwards, the multi-dimensional features of filtered signals are classified using the designed DF model, which could deeply express the difference between radar targets and clutter. Finally, comparison has been conducted against CFAR and CNN by using experimental data and the results indicate that the proposed method significantly improves radar target detection capability under complex environment.

The information conveyed through facial expressions accounts for a large proportion of the total information and can effectively express people's intentions and emotions. Facial expression recognition has laid the foundation for fields such as human-computer interaction, facial emotion prediction, and artificial intelligence, and has become an important research object in computer vision. This article proposes a facial expression recognition method based on the MobileNetV3 network for face images from different angles. The method uses depth-wise separable convolution, introduces attention mechanism and new activation function to update blocks, and redesigns the time-consuming layer structure at the end. The dataset used in this article is the KDEF, which includes 4,900 color images with a size of 562*762 pixels. Through extensive experiments, it has been shown that the proposed structure in this article improves the accuracy of facial expression recognition from different angles compared to other network structures, achieving 94.7%, and has a smaller parameter count, which is beneficial for further research on facial expressions.

To solve the problems of intra-class difference and modal difference in cross-modality person re-identification tasks, person re-identification based on two-layer attention and joint learning is proposed. Firstly, the input visible images are enhanced by gray-like channels for joint learning. Then, to obtain more effective feature semantics in space and channels, a two-layer attention module is embedded behind the third convolutional layer of Resnet50. Finally, to promote network convergence and reduce the difference between different modalities, a modality distribution difference loss function is utilized. The effectiveness of this algorithm has been experimentally proved not only in two mainstream public datasets SYSU-MM01 and RegDB, but also in the newly proposed dataset LLCM. The effectiveness of these three modules is proved by ablation experiments on the SYSU-MM01 dataset.

In response to the issues of limited annotated data and insufficient semantic information extraction in medical named entity recognition for Chinese electronic medical records, a novel approach based on data augmentation and structural feature fusion is proposed. Firstly, addressing the potential noise introduced by unsupervised data augmentation methods like synonym replacement and random insertion, a rule-based data augmentation technique is introduced. This method effectively diversifies the annotated Chinese electronic medical record data by making use of external resources. Considering that Chinese characters are ideograms and characters with similar radical structures often convey similar semantic information, a semantic information enhancer is proposed. This enhancer utilizes a convolutional neural network to obtain vector representations of Chinese character features. Based on cosine similarity, it extracts several characters with the highest radical structure similarity, and incorporates additional semantic information through an attention-weighted fusion mechanism. This enables the model to extract semantic information more effectively. Lastly, a semantic fusion module is constructed to integrate the semantic information enhancer into the primary named entity recognition model, employing a Conditional Random Field (CRF) to predict the final results. Experimental results on the CCKS 2019 Chinese electronic medical record dataset demonstrate that the model achieves an impressive F1 score of 83.71%, validating the effectiveness of this approach.

With the continuous popularization of artificial intelligence, the establishment of transmission line image automatic detection system is imminent. Due to the small number of samples and high similarity of transmission line image data sets, traditional image detection algorithms cannot achieve real-time positioning and accurate identification of power equipment components. In view of the above problems, this paper proposes an intelligent recognition method of transmission line defect edge based on deep learning. Firstly, the convolutional neural network (CNN) is used to extract the feature of transmission line defect graphics, and a large number of sample data are obtained. Secondly, the edge algorithm is used to optimize the CNN algorithm to determine the pixels, and the transmission line defect detection is completed. Finally, it is applied to the self-built transmission line data set for verification. The average detection accuracy of this method is 89.2%, which has excellent detection accuracy and generalization ability.

Aiming at technical advantages of quickly discover and real-time tracking focused on targets with UAV video, we propose a multi-object tracking method based on spatial constraints. Utilizing the pre-training model of YOLOv7, we do a little work of model modification with UAVDT dataset and a small self-made dataset, then train a special object detection model for identifying and positioning vehicles in the battlefield environment; according to features of battlefield environment, we adopt classification and grading association method to determine the correlation between detection box and object, and propose a data association method based on spatial constraints and object re-identification, compared to current popular algorithms, which effectively improved the ID_SW and reduce the number of object ID switching.

Nowadays there is a problem of "semantic gap" in computer vision. The existing text-based caption can not fully utilize the textual information in the image and the text recognition step may generate false recognition results, but there is no modification mechanism to correct the errors. In this work, we improve the existing approach by proposing a text-aware recognizer to extract image text information from the input data and generate corresponding text descriptions and text features. Considering the relationship between the text object and the image content, in order to improve the semantic errors in the text description sentences generated by the character recognizer, we introduce the caption-rectify module, which can better improve the text information involved in the image and model the text information recognized in the textcaps dataset. Seriously speaking, we propose to use the current state-of-the-art text recognizer to detect characters and generate contextual descriptions of images. Moreover, we propose a correction mechanism and demonstrate qualitatively and quantitatively that the correction can make the final caption statement consistent with the textual information in the image, improving the semantic accuracy of the text description. We validated our approach on text caption task, thoroughly analyzed each module, and showed significant improvements compared with the current advanced model LSTM-R and CNMT.

With the advancement of deep learning frameworks, object detection algorithms based on deep learning have gained widespread adoption across various domains, including mechanical manufacturing, urban transportation, and medicine. Compared to traditional vision-based template matching methods, deep learning-based object detection methods exhibit lower requirements for object characteristics, superior generalization capabilities, and higher accuracy in detection. They are particularly suitable for detecting similar objects against complex backgrounds. This paper provides a comprehensive review of two-stage and one-stage object detection network architectures by summarizing their advantages and disadvantages. Additionally, it discusses the backbone networks used, training datasets employed, evaluation metrics utilized for comparing target detection algorithms' accuracy and response speed. Furthermore, it presents improved approaches for enhancing target detection algorithms' performance. Finally, this study proposes research focus areas and future development directions in the field of object detection.

In order to improve the adaptability of unmanned wave glider to complex sea states, a reduced-order extended-state-observer-based line-of-sight (ELOS) guidance law is designed for the adaptive path tracking control problem of UWG. In this ELOS guidance law, the reduced-order extended state observer is introduced to estimate the time-varying sideslip angle, so that the unknown time-varying sideslip angle can be compensated, which improves the control performance of the system and accelerates the convergence speed. The stability analysis and simulation comparison experiments verify that the control method designed in this paper has good robustness and a significant control effect.

Safety helmets are the most common and practical personal protection tool on construction sites and can effectively prevent and mitigate head injuries caused by accidents. Detecting safety helmets and various safety equipment is a major task in the safety management of construction site personnel. It is also an important element in the intelligent monitoring technology of construction sites. With the development of artificial intelligence, it has now become an important part of intelligent construction site construction. The safety equipment target detection algorithm is investigated to comprehensively analyze the research status of artificial intelligence technology in safety equipment detection. In this paper, after introducing the attention mechanism into the yolov5s algorithm and adding the density residual module RDB and the cavity density residual module DRDB proposed in this paper, the accuracy of identifying helmet wearers is improved by 1.4% for aerial scenes and about 1.2% for seat belt wearers, and the accuracy of identifying helmet wearers is also improved slightly for ground scenes.

The clinical diagnosis of gastrointestinal diseases heavily relies on endoscopic examinations, utilizing medical electronic endoscopy systems that comprise various components. Gastrointestinal endoscopy categorizes regions, with upper and lower digestive tracts distinguished. The manual diagnosis process encounters challenge due to factors like variability, time constraints, and extensive image pools, leading to missed diagnoses. This paper proposes an automated bubble analysis and recognition method using Transformers, addressing image quality issues caused by motion blur, excessive brightness, and inadequate exposure. A Transformer-based neural network model is constructed to classify and diagnose bubble images. Contributions include a neural network model, targeted preprocessing techniques, and a performance evaluation with a substantial collection of gastrointestinal bubble images, demonstrating high accuracy in classification and diagnosis for both clear and blurry images. On clear images, the TOP1 accuracy can reach 0.932, and on blurry images, the TOP1 accuracy can reach 0.9375.

Due to the advantages of high throughput, low latency, and low power consumption, optical neural networks hold great promise in addressing the challenges of energy consumption and computational efficiency faced by current artificial intelligence technologies. Consequently, they have become a research hotspot in both academia and industry in recent years. The goal of optical neural networks is to use photons as the physical carrier to construct the basic computational units of artificial neural network algorithms, thus achieving high-performance novel computing architectures and applying them to solve practical problems. This paper introduces the working principles and characteristics of optical neural networks and discusses relevant research on target recognition based on optical neural network architectures.