Small target detection algorithm for UAV aerial photography based on attention mechanism

Low detection accuracy and high miss detection rate are mostly caused by target blur, occlusion, and complicated environment from the drone's point of view. We propose a detection algorithm based on the attention mechanism by improving YOLOv8s. The dynamic sparse attention mechanism BiFormre is used in the backbone network to improve the global awareness ability of the model, allowing the model to make full use of contextual information to assist object detection. Scale-awareness, space-awareness, and task-awareness attention mechanisms are used in the detection head of the Dy Head dynamic attention mechanism detection head to perform multi-scale object detection. The issue of poor small target recognition performance brought on by abrupt changes in target scale can be successfully resolved by including a small target detection layer in the feature fusion stage. Experimental results show that we achieved 36.9% mAP@0.5 on VisDrone-2019DET, which is 5.7% higher than the benchmark model YOLOv8s, and still performs better than other popular algorithms.