Research On Key Issues In Fault Detection Of Rail Fastener Based On Image

With the increasing of China’s railway mileage,especially high-speed railway,the track inspection work to ensure the safe operation of high-speed railways is becoming more and more onerous.Considering the safety of track inspection workers and to reduce their working intensity,China has developed high-speed comprehensive inspection train(HSCIT)that can help inspect tracks.HSCIT is equipped with numerous cameras to collect track images for automatic fault detection of track facilities.Due to the vibration of HSCIT and the limitations of camera,the collected track images will inevitably be contaminated by noise during the acquisition,which will affect the subsequent fault detection.The thread running through this thesis is the influence of noise on image fault detection of rail fasteners.We focus on the unknown noise mode,the difficulty of distinguishing noise and image features,and the number of fault samples is far less than the number of normal samples in the fastener images collected by HSCIT.The research on these issues is as follows:(1)We establish a noise model of the images collected by the HSCIT and study a noise level estimation algorithm.For the noise model,according to the similarity between rail fastener images,we exploit the probability distribution of the histogram to approximate the probability distribution of noise as region adaptive Gaussian noise,and verify the correctness of the noise model hypothesis with quantile-quantile plot.To estimate the noise level in the image,we select the low-rank regions in the image and calculate the eigenvalues of the covariance matrix of the selected regions.With these eigenvalues as the initial data set,we continuously compare the relationship between the median and the mean of the eigenvalues in the data set to eliminate the excessive eigenvalues.A suitable number of eigenvalues are selected for noise estimation.Experiments show that the noise level estimated by this algorithm is higher than that of all comparison algorithms.And in image denoising,the higher the estimated noise level is,the better the denoising effect of color image is.(2)Aiming at the established adaptive noise model,we propose a region-adaptive weighted nuclear norm minimization(RA-WNNM)model for gray image denoising.Based on the existing Gaussian denoising model,RA-WNNM embedded the feature of noise adaption with region into the optimization function in the form of weight matrix.We derive the values of the weight matrix by using the maximum posterior probability.There is no analytical solution for RA-WNNM.We transform RA-WNNM from an unconstrained optimization problem to an equal-constrained optimization problem,which can be decomposed into several subproblems with analytical solutions iteratively by using alternate direction multiplier method.Taking into account the noise difference between different channels of color images in reality,we extend RA-WNNM to color image denoising.Experiments demonstrate the effectiveness of RA-WNNM in rail fastener images denoising.(3)In view of the phenomenon that the faulty images of the fastener are far less than the normal images,we conduct research in two different directions.One direction is the image classification model of low-rank representation and shared dictionary learning(LRRSDL).The other direction is the few shot learning model based on transfer learning.LRRSDL considers that there are similarities among image samples of the same class,especially among fastener images.LRRSDL adds the similarity of coded coefficients between similar samples as regularization term in the form of nuclear norm into the objective function.Experimental results show that LRRSDL performs well on rail fastener data set.And we verify the negative effect of noise on image classification through experiments.In the second direction,we adopt transfer learning to tackle the problem of small samples.We utilize graph and convolutional neural network to extract semantic knowledge and image features from large image classification data sets.Then we inherit some of the learned network layers and fine-tune network parameters through small sample data sets.For the railway fastener data set,the classification performance of this network is slightly inferior to LRRSDL.