YOLO-RS: A More Accurate and Faster Object Detection Method for Remote Sensing Images

In recent years, object detection based on deep learning has been widely applied and developed. When using methods to process remote sensing images, the trade-off between speed accuracy of models is necessary, because images pose additional difficulties such as complex backgrounds, small objects, dense distribution task. This paper proposes YOLO-RS, an optimized algorithm YOLOv4 address challenges. The Adaptively Spatial Feature Fusion (ASFF) structure introduced after feature enhancement network YOLOv4. It assigns adaptive weight parameters fuse multi-scale information, improving accuracy. Furthermore, optimizations are Pyramid Pooling (SPP) in By incorporating residual connections employing 1 × convolutions maximum pooling, both computation complexity improved. To enhance speed, Lightnet introduced, inspired by Depthwise Separable Convolution for reducing model complexity. Additionally, loss function introducing Intersection over Union function. change replaces aspect ratio term with edge length loss, enhancing sensitivity width height, accelerating convergence, regression detected frames. mean Average Precision (mAP) values YOLO-RS 87.73% 92.81% under TGRS-HRRSD dataset RSOD dataset, respectively, which experimentally verified be 2.15% 1.66% higher compared original algorithm. reached 43.45 FPS 43.68 FPS, 5.29 Frames Per Second (FPS) 5.30 improvement.