UAV image defect detection method for steel structure of high-speed railway bridge girder

Railway bridge is an important infrastructure along the railway, which needs regular inspection and maintenance. Using UAV technology to inspect railway infrastructure is a hot issue in current research. However, because of the detection target of steel structure of railway bridge girder, such as bolts, rust, etc., are small and dense targets, and the manual annotation workload is huge in the early stage. Therefore, this paper first proposes a method for constructing the sample set of semi-automatic annotation of railway bridge defects, which can greatly improve the efficiency of data annotation. Secondly, in order to solve the problem that it is difficult to detect large-scale images of UAV and the target scale changes dramatically due to the flight distance and height changes, this paper proposes an adaptive cropping strategy for image of steel structure UAV of railway bridge girder. Compared with YOLOv5s that does not adopt the strategy, its mAP improves by 32%. In order to reduce the GPU memory usage and the number of parameters, and improve the parallel running efficiency of the model, the GYOLOv5 model is proposed by combining Ghost Bottleneck and SIoU loss function. Compared with YOLOv5s, which also adopts the adaptive cropping strategy, the mAP of GYOLOv5 model increases by 6%, and the number of parameters can be reduced by 1.7 million. Finally, in order to help the network to find the region of interest in the images covered by the UAV and improve the detection accuracy, the Attention mechanism of CBAM (Convolutional Block Attention Module) is combined with YOLOv5s. Compared with the original YOLOv5s model, the number of parameters is only 600, but the total mAP is increased by 12%.