A deep learning-based optimization strategy for small object detection

In high-altitude images captured by UAVs, small targets often appear densely packed, occupy fewer pixels, and suffer from mutual occlusion, making accurate detection a significant challenge. To address this issue, this paper proposes an optimization strategy based on YOLOv5s. First, the DHDCNv4 detection head is introduced to enhance the feature detection capability across different scales of the corresponding feature layers by incorporating unified attention mechanisms. Second, to reduce the sensitivity of IOU to small target displacement, the NWD loss function is employed as a replacement for the traditional IOU. Finally, the proposed approach is evaluated on the Visdrone2019 dataset. Experimental results demonstrate that the proposed method outperforms the original YOLOv5s in detecting small targets in high-altitude UAV imagery.