An Optimized YOLOv5s-rd Framework for Efficient Target Detection in Remote Sensing Images

Remote sensing image object detection methods are crucial for applications related to aircraft, ships, vehicles, and buildings. Traditional methods, relying on manually designed features, suffer from high computational complexity, leading to low detection efficiency and stability. In response, we present an enhanced remote sensing image object detection approach, YOLOv5s - rd, which is built upon an optimized YOLOv5s.Our method integrates structural enhancements, refined loss functions, and advanced data augmentation strategies. These improvements include optimizing the number and orientation of rotation anchors to better handle target scale diversity and rotation changes in remote sensing images. We also adjust Gaussian distribution parameters, which is beneficial for dealing with the challenges of complex backgrounds. Additionally, we calibrate the weights of the weak - supervision branch, considering the fact that the number of objects in remote sensing images is usually small, aiming to improve the model's performance with limited labeled data. We conduct experiments on multiple public datasets: DOTA, HRSC2016, UCAS - AOD, and Northwest University VHR - 10. The results demonstrate that YOLOv5s - rd outperforms traditional and existing deep - learning methods in detection performance. Specifically, on the DOTA dataset, it achieves a mean average precision (mAP) of 80.4% and 41.2 FPS; on the UCAS - AOD dataset, 96.7% mAP and 40.3 FPS; on the HRSC2016 dataset, 93.2% mAP and 38.7 FPS; and on the Northwest University VHR - 10 dataset, 95.2% mAP and 39.7 FPS. Moreover, its computational complexity (FLOPs) is only 11.0B, surpassing most of the compared methods. By combining these novel optimizations, our YOLOv5s - rd not only enhances the robustness and effectiveness of the detection model but also significantly improves performance and reliability compared to existing methods, providing a new solution for remote sensing image object detection.