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