The Research On Visual Detection Method For High-Speed Railway Surface Defect Based On Deep Learning

High-speed railway is the backbone of comprehensive transportation system in China,and quality of track is related to the safety and stability of train operation,so that it is no doubt about the regular inspection and long-term maintenance.However,traditional manual inspection and contact measurement techniques are still used to monitor the fault damage,which exists wrong or missed judgment,low efficiency,and subjective assumption,which cannot meet the command of recent advanced railways.The existing technologies faced several challenges such as early detection,reliability and system cost.It is most practical to study an automatic visual detection technology for rail surface under current railway line construction.In view of this,this paper makes an in-depth and systematic research on rail surface defect detection based on deep learning.Firstly,it expounds the national strategy of high-speed railway construction and development;The importance and necessity of developing visual detection of railway track surface defects are pointed out.This paper introduces the research status and related literature at home and abroad,and analyzes the engineering problems and challenges of high-speed railway surface defect detection.The main work is as follows:1.According to the task requirements and working environment,we build a machine vision-based hardware for track detection,and especially discuss the performance indexes like illumination,imaging scheme,view field of camera and optical lens,which is the foundation of image acquisition for detection methods in our system.2.A multi-mode parallel fusion defect detection method is proposed where an improved Gaussian mixture model based on Markov random field FRGMM incorporates spatial correlation for fast and accurate pre-segmentation,and the expectation maximum(EM)algorithm satisfies the direct solution of posterior distribution.To exclude non-defect,the advanced Faster R-CNN applies k-means algorithm for dimension cluster to learn several labeled features under multiple background such as uneven reflection,noise,corrosion and oil stains,to capture the real defect location.Finally,the joint area is hit by global segmentation proposal and local target position,which is a real defect.3.A multi-stage cascaded defect detection method is proposed to solve the orbital reflectance,et al.Technical route includes ROI extraction,background equalization,denoising,and segmentation.Firstly,ROI extraction algorithm based on gray vertical projection is proposed to obtain rail image of high resolution,and regularization term of variational model is used as curvature filter for image denoising.Finally,a segmentation algorithm based on residual pyramid pooling network(RPPNet)and Legendre level-set is presented to obtain the contour of surface defects from coarse to fine within a FCN framework.4.Bayesian CNN and attention mechanism-based defect detection method called as Deep Rail is proposed.Relying on Deeplab v3+ skeleton network Xception,the Dropout layer is added to build the probability model so that Monte Carlo samples are generated from posterior distribution.Atrous spatial pyramid pooling ASPP extends the receptive field,and decoder refines the object boundary and outputs Softmax probabilistic mean and variance as segmentation and confidence.To solve the long-tail problem,attention mechanism is introduced to balance the loss of foreground and background for optimal weight.