An Efficient Small Target Detection Algorithm for UAV

Detecting small targets in drone imagery is a critical task across various domains. However, current detection algorithms show limited sensitivity to highly overlapping small targets in densely occluded scenes, resulting in frequent false positives and missed detections under complex conditions. To overcome these limitations, we enhance YOLOv8s with two key modifications: (1) reducing complicated background interference by incorporating a Large Selective Kernel (LSK) attention mechanism into the backbone network, and (2) replacing the standard detection head with a DyHead module. This architecture dynamically fuses multi-scale features using adaptive attention mechanisms, significantly improving multi-target discrimination. Evaluations on the VisDrone2019 dataset show that our method achieves a 4.2% increase in mAP@50 and a 2.7% improvement in mAP@50:95 compared to the baseline. Experimental results further demonstrate the efficiency of the strategy in boosting small target detection performance for drone applications.