Research On Detection Algorithm For Rail Surface Defect Based On Cascade R-CNN

As an important part of railway,rail is prone to appear surface defect and internal defect under long-term repeated load,and the main performance is surface defect.Rail surface defect will not only make the train generate vibration,noise,affect the quality of the train operation,but also when the defect to a certain extent,will lead to rail fracture,even derailment,overturn and other accidents,affect the safety of the train operation.Therefore,it is of great significance to research rail surface defect detection.Among the commonly used rail surface damage detection methods,physical detection methods such as eddy current detection,magnetic flux leakage detection,laser ultrasonic detection,acoustic emission detection,and traditional image processing methods cannot achieve good detection results.Therefore,this article researches how to use deep learning-based rail surface defect detection method to realize accurate and rapid detection of rail surface defect.The main work is as follows:1)In order to solve the problem that insufficient samples of rail surface defect images could easily lead to over-fitting in the training process,the dataset of rail surface defect images was expanded by using the data enhancement methods,such as flipping transformation,random clipping,brightness transformation and generative adversarial networks,so as to increase the number of rail surface damage images.2)In order to realize the accurate detection of rail surface defect with small size characteristics,Cascade R-CNN algorithm was applied to rail surface damage detection,and its improvement was made.Firstly,IOU balanced sampling method was adopted to improve the sampling method of Cascade R-CNN on candidate region,and the proportion of difficult samples in the training samples obtained by sampling was increased.Secondly,region of interest align(Ro IAlign)is adopted to solve the problem of misalignment between candidate region and extracted feature map caused by two rounding quantization in region of interest pooling(Ro I Pooling).Finally,Complete Intersection over Union(CIOU)loss is used to solve the problem of inaccurate regression of Smooth L1 loss.The experimental results show that the three improved methods used in this paper can effectively improve the detection accuracy of rail surface defect.3)In order to reduce the amount of calculation of the detection network and further improve the speed of rail surface damage detection,this paper researches the use of lightweight networks Mobile Net V3-Large and Shuffle Net V2 to improve the feature extraction network in Cascade R-CNN,and adaptive spatial feature fusion(ASFF)method is used to improve the feature map fusion module,and then two lightweight rail surface damage detection networks,Mobile-R-CNN and Shuf-R-CNN,are constructed.The experimental results show that both Mobile-R-CNN and Shuf-R-CNN improve the detection speed on the premise of ensuring a certain detection accuracy,and Mobile-R-CNN is superior to Shuf-R-CNN in both detection accuracy and detection speed.In this paper,the rail surface defect detection is researched from two angles of detection accuracy and detection speed,and the accurate and rapid detection of the rail surface damage is realized.It can provide technical support for the rail surface defect detection,and has high theoretical and practical value.