Small Target Detection Algorithm for UAV Aerial Images Based on YOLOv8n

This paper proposes an improved YOLOv8n algorithm to overcome the limitations of existing detection methods in UAV aerial imagery, particularly focusing on the issues of small target size and significant occlusion that typically degrade detection performance. The enhanced algorithm is specifically designed to boost detection accuracy for small targets in drone-captured images. Firstly, a deformable convolutional module C2f_DCNv2 is designed to integrate deformable convolutions to enhance the ability of features extraction. Secondly, to enhance multi-scale feature representation, we reconstruct the feature fusion network by incorporating both a dedicated tiny target detection layer and BiFPN structure. This design enables more comprehensive cross-scale feature integration while better preserving fine-grained details crucial for small object detection. Thirdly, the ITA attention mechanism is refined by leveraging the triplet attention mechanism, enabling the model to enhance its focus on regions of interest and suppress distracting background information. Finally, Wise-IoU replaces CIoU as a regression loss function to improve the positioning accuracy of the detection frame. Experiments on the VisDrone2019 dataset show that the mAP50 value of the improved algorithm reaches 40.7%, which is 7.9% higher than that of the baseline model YOLOv8n, and has better detection performance than other mainstream target detection algorithms.