Rail Surface Defect Detection Based on An Improved YOLOv5s

As the operational time of railway increases, rail surfaces undergo irreversible defects. Once defects occur, it is easy for them to develop rapidly, which seriously threatens safe operation trains. Therefore, accurate and rapid detection surface very important. However, in defects, there are problems, such as low contrast between background, large scale differences, insufficient training samples. we propose a defect method based on an improved YOLOv5s this paper. Firstly, sample dataset images was augmented with flip transformations, random cropping, brightness transformations. Next, Conv2D Dilated Convolution(CDConv) module designed reduce amount network computation. In addition, Swin Transformer combined Backbone Neck ends improve C3 original network. Then, global attention mechanism (GAM) introduced into PANet form new prediction head, namely transformer GAM Prediction Head (SGPH). Finally, used Soft-SIoUNMS loss replace CIoU loss, accelerates convergence speed algorithm reduces regression errors. The experimental results show that reaches 96.9% average precision detection, offering has certain engineering application value.