Research on UAV Aerial Image Detection Model Based on CBC-YOLOv10

Aiming at the problems of scene complexity, multi-target aggregation, and mutual occlusion in aerial photography small target detection, this paper proposes the CBC-YOLOv10 detection model. The residual block in the C2f module of the model backbone network is improved to form C2f-CAS, which combines the convolution operation with the additive self-attention mechanism to effectively reduce the computational complexity. In the neck network to solve the problem of redundancy of feature information or unreasonable weight allocation, BiFPN-MAF is introduced, which realizes the fusion and information transfer of features of different scales and retains more feature details. C3-DCNv2 is used instead of the traditional node module to improve the detection accuracy of irregularly shaped targets. Validated on the VisDrone2019 dataset, the parameter amount is reduced by 19.08%, [email protected] and [email protected]:0.95, which are improved by 3.9% and 2.7%, respectively.