UAV-YOLOv8n: An Improved YOLOv8 to Detect Small Target in UAV Remote Sensing Images

UAV target detection is widely used in commercial, civil and military fields. However, UAV aerial images have few small target pixels and a large number of small targets, so there are omissions and false detections in target detection. Based on the above problems, an advanced YOLOv8n aerial photography detection algorithm is proposed. First, the RepNCSPELAN4 module is introduced, which has the functions of enhanced region representation and multi-scale fusion, which can effectively improve the correctness and robustness of the detection algorithm. Second, the SimAM (Simple Parameter-free Attention Module) attention mechanism is added after the backbone module and the RepNCSPELAN4 module to focus on the target region and refine the feature information. In addition, the convolution modules in the first and third layers of the backbone layer are replaced with dynamic convolution, which has higher parameter efficiency and feature expression ability, and can effectively improve the small target detection ability. Experiments show that the mAP50 of the proposed model is improved by $1.8 \%$ on the VisDrone2021 public dataset, respectively, and there is also a significant improvement on the SeaDronesSee dataset.