Abstract: Aiming at the problem of safety hazards of insufficient bridge crack detection, we  proposed a bridge crack detection algorithm based on YOLOx-s, combined with a small unmanned aerial vehicle platform. Firstly, we added a residual hole convolution module in  the backbone to solve the problem of   large scale changes and complex backgrounds in drone images. Secondly, we added a coordinate attention mechanism module in  PANET to improve the detection rate of small targets. Finally, we replaced the loss function with Focal loss to enhance the learning of positive samples and improve the stability of the model.  The experimental results show that compared with the YOLOx-s algorithm, the proposed method improves detection accuracy by 3.72 percentage points. On embedded devices, this method has better accuracy than other mainstream algorithms and can achieve real-time detection, which can be better applied in bridge crack detection for unmanned aerial vehicle.

Key words: unmanned aerial vehicle (UAV), bridge crack detection, object detection, YOLOx-s algorithm, attention mechanism