Research On Track Defect Detection Algorithm Based On Machine Vision

Railway security is a necessary section of railway operation guarantee system.With the continuous advancement of railway planning and railway network construction,as well as the continuous expand of railway operation time,the track defects such as track surface defects and abnormal fasteners are gradually prominent.Therefore,timely and accurate detection of these defects is of great significance to ensure the security of railway operation.Traditional detection algorithm is complex and lack of generalization ability,which cannot satisfy the current different railway surroundings and application requirements of detection scenarios.Compared with the usual method,the detection speed and precision of track defect based on machine vision are radically improved.Moreover,it can learn the target characteristics independently according to the algorithm,and has few restrictions on the environment,so as to realize the accurate detection of track defects.In this dissertation,the detection models for rail surface defects and abnormal fastener detection is proposed.Meanwhile,the two models are improved based on railway state to meet the requirements of railway application.For the defect problem of rail surface,the U-Net model is used as the defect segmentation algorithm.(1)Residual structure is used to substitute the convolution layer.Not only can the model depth be increased,but also the features extracted from the track surface defects can be added to compensate for the loss of edge detail caused by pooling,thereby making the segmentation more accurate.Moreover,the residual structure uses identity mapping to send shallow information to the deep network,which improves the precision of model segmentation.Meanwhile,it avoids problems such as the disappearance of gradients caused by the increase in network depth.(2)Add Large kernel matter structure to skip structure.By adding large kernel matter into the skip structure,not only can the global information extraction be improved,more detailed information of orbit surface defects can be extracted,but also the features of high and low layers can be better fused.At the same time,the noise introduced by the shallow layer can be reduced and the overall segmentation precision of the model can be enhanced.(3)Introducing attention mechanisms into the model.In the original model structure,prediction and segmentation are generally carried out on the whole image,so the network will extract excessive background features and reduce the segmentation effect on the defect position on the rail surface.By introducing the attention mechanism,not only can the weight of the defect part be increased.And the background and noise can be suppressed,so that the model can reduce the parameters and improve the segmentation accuracy of the defect.For fastener abnormal problems,Faster R-CNN is used as fastener detection algorithm.(1)multi-layer fusion.Resnet network is used to replace the original feature extraction network,and the number of residual blocks can be set flexibly according to the requirements,so that more details and spatial information of images can be extracted,which makes model classification and positioning more accurate.In feature extraction,multi-layer fusion technology is used to integrate low-level spatial information into the feature map,which makes the model more accurate for the location of fastener.(2)Reduce the number of proposals in RPN network.In the original network,it is difficult to predict the size and position of the target,so a large number of different specifications and proportions are used to search the whole image.However,in the field of fastener detection,the image position of the target has little change,and it can only be considered to use proposals of different sizes to predict it,which can reduce the number of proposals and improve the detection speed of the model.In order to measure the overall performance of the two models in the detection of track defects,the models are compared with other models in the same field.In rail surface defect segmentation and fastener abnormal detection,the detection accuracy and speed of the algorithms proposed have been improved.The segmentation accuracy of rail surface defects reaches 80%,and the average detection accuracy of abnormal fastener reaches 94%,both of which are superior to the existing detection algorithms in the research of railway defects.