Detecting Small Object in RSI by Improved Yolov5 and Faster R-CNN Models

Focusing on the small, densely arranged objects and complex background areas in remote sensing image (RSI) object detection, improvements are proposed to both the single-stage object detection model, Yolov5, and the two-stage object detection model, Faster R-CNN. The aim is to improve the accuracy and speed of the single-stage model and enhance the precision of the two-stage model. For the Yolo model, the Shuffle Attention mechanism is introduced. This mechanism groups channel features, applies spatial and cross-channel attention, and fused sub-features to enhance detection accuracy and speed. In the Faster R-CNN section, the VGG16 backbone network is replaced with ResNet50 featuring skip connections, which mitigates gradient issues and supports the training of deeper networks. Experimental results on the DOTA dataset, compared to those of the pre-improvement models, indicate that the improved Yolov5 reduces processing time by 0.4 hours, increases the overall average precision by 0.3%. The improved Faster R-CNN exhibits significant improvements in parameters such as average precision, mean average precision, and recall.