Improved YOLOv8n Algorithm for Small Target Detection in UAV Perspective

Aiming at the problem that the target is small and easily obscured when detecting target UAVs in UAV aerial images, an improved target detection algorithm based on YOLOv8n optimization (RDC-YOLOv8n) is proposed. First, the RepViT module is introduced into the backbone network and the RepViT_DAM module is designed to replace the original Conv module. The RepViT module improves the inference efficiency and reduces the computational complexity through the structural reparametrization technique. The RepViT_DAM module improves the robustness of the model through the enhancement of the fusion of contextual information. Secondly, the C2f_concat module is designed to replace the original C2f module, which enhances the feature processing capability of the backbone network and reduces the misdetection and omission rate of small targets in UAV images by increasing the branches that obtain information directly from shallow features. Experimental results show that the improved RDC-YOLOv8n algorithm is tested on the public dataset MOT_FLY_V0.3, which shows that the mAP50 reaches 62.8% and the mAP50:95 is 28.1%. This is an improvement of 6.3% and 2.3% over YOLOv8n, respectively. Meanwhile, the number of parameters of the model is kept at a low level, which is suitable for real-time target detection on resource-constrained devices such as UAVs.