Proceedings Article | 21 June 2024
KEYWORDS: Object detection, Detection and tracking algorithms, Convolution, Remote sensing, Network architectures, Feature extraction, Data modeling, RGB color model, Performance modeling, Design
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.
