Improvement of Small Target Detection Algorithm Based on YOLOV5

Target detection has always been a difficult problem in computer vision. The commonly used target detection algorithms are Two stage and One stage. In order to address small ground targets and reduce the storage of UAV, the paper present an improved method on the basis of YOLOV5. All C3 modules are replaced by in the backbone network. It can improve the speed and accuracy of the network target detection, and improve the efficiency of the model training and inference. A new feature fusion mechanism is proposed, which greatly improves the feature acquisition and fusion capabilities for networks. A 160*160 detection head is added at the end of the network. Experimental Specific improvements resulted in a 3.6% increase in recall, a 3.9% increase in mAP_0.5, and a 2.8% increase in mAP_0.5:0.95.