UAV aerial photography target detection based on improved YOLOv9

UAV aerial photography target detection based on improved YOLOv9

연구 분야: Verification

학회: The Journal of Supercomputing

In UAV aerial photography, the existence of small-size targets, dense distribution and occlusion phenomenon often leads to frequent missed and false detection in the detection process, which has a significant impact on the detection accuracy of the model. To solve this problem, this paper proposes an improved YOLOv9s model, BF-YOLOv9s. First, the application of the BiFormer attention mechanism serves to enhance the model’s concentration on small targets, thereby facilitating the retention of more refined and detailed features. Second, according to the lightweight demand of UAV aerial photography, the RepNCSPELAN4_Ghost module is proposed, which integrates GhostConv into the backbone network RepNCSPELAN4, significantly reducing the computing load and optimizing the use of computing and memory resources. Finally, the BiFPN feature pyramid network is introduced to promote the fusion and exchange of cross-layer information and improve the detection effect. By selecting the Focal WIOU loss function, model convergence is accelerated, the loss is reduced and training efficiency is improved. The experimental results show that BF-YOLOv9s achieves a mAP50 of 41.3% on the VisDrone2019 dataset, outperforming the original YOLOv9s by 5.6%, while also reducing the parameter count by 8.3%.