DPN-YOLO: an enhanced algorithm for small target detection in UAV imagery
摘要
To mitigate the persistent challenges of occlusion and scale variation in small target detection within UAV imagery, this study proposes an enhanced YOLO11-based algorithm incorporating several key innovations. These include the implementation of a dedicated small target detection head to improve sensitivity, the adoption of the SCDown method for efficient feature retention during downsampling, the introduction of a novel C3k2_PPA module with multi-branch fusion and attention mechanisms to enhance feature representation, and the replacement of the CIoU loss function with the NWD loss. Evaluated on the VisDrone2019 dataset, the proposed method achieves an accuracy of 48.2% and a recall of 35.2%, representing improvements of 7.2% and 5.2% over the baseline YOLO11n, respectively. Furthermore, it attains mAP50 and mAP50-95 scores of 35.7% and 20.5%, corresponding to gains of 6.4% and 3.8%. These results demonstrate that the improved algorithm offers enhanced robustness and accuracy compared to the original YOLO11.