Research On Classification Algorithm Of Railway Fastener Combined With Structural Information

Railway fasteners are the connecting parts of rails and sleepers,which are important guarantees for railway transportation safety.It is of great engineering value to detect the conditions of fasteners by machine vision technology.The fastener detection tasks belong to the image classification problem.After obtaining the fastener images through the track inspection vehicle,the fasteners are classified into normal category and invalid category according to the low-level features of the images,which are the commonly used fastener detection algorithms.However,due to the factors such as different forms of the fasteners,obstructions of railway ballast,the change of illumination,and the various forms of failure,the low-level features cannot describe the content of the images stably,resulting in that the classification results are different from the true semantics of the fastener images.The semantic gap between the low-level features and the real content of the images limits the accuracy of the fastener detection algorithm.This paper designed a robust feature semantic learning method that can express the structure of fasteners.The main studies of this paper are as follows:(1)A texture feature extraction algorithm combined with noise estimation(AT_NRLBP).Aiming at the problems that the traditional Local Binary Pattern algorithm is sensitive to image noise and the fixed threshold of Noise-Resistant LBP algorithm cannot correct part of the image noise.This paper proposed an adaptive threshold Noise-Resistant LBP algorithm combined with the noise estimation algorithm,which named AT_NRLBP.In this paper,the correlation between the noise intensity of image and the coding threshold of Noise-Resistant LBP was analyzed.According to the AT_NRLBP algorithm,the images were evenly divided into sub-blocks first.In addition,the noise levels of the image sub-blocks were estimated.Finally,the threshold of Noise-Resistant LBP was determined by the noise intensity of the image sub-block,which enabled adaptive selection of the encoding threshold.Theoretical analyses show that AT_NRLBP algorithm determines the coding threshold by local noise estimation of image sub-blocks,which can eliminate the influence of global noise and reduce the influence of image noise on texture feature coding.AT_NRLBP algorithm can extract texture features of images more accurately.The experimental results show that the classification accuracy of AT_NRLBP on the KTH-TIPS database is 2.9% higher than that of NRLBP.(2)A fastener classification model combining global and local constraints(glc-sLDA).Aiming at the problem that the test images lack manual labels in supervised LDA model,resulting in that the topic distributions of test images ignored the structure information.This paper proposed a fastener image classification model combining global and local constraints based on supervised LDA,which named glc-sLDA.The global constraints described the class states of the test images,and the local constraints expressed the topic similarities between the test image sub-blocks and the training image sub-blocks.To begin with,the training topic distributions with structural information were obtained by sLDA model.And then,the products of the global constraints and the local constraints were calculated,which were the topic update weights.Finally,the test topic distributions were updated by weighted summation of training topic distributions.Theoretical analyses show that the structure information of test image topic distributions were took into account in glc-sLDA model,which could reflect the structure states of fasteners more accurately.The experimental results show that the glc-sLDA model can express the structure information of fasteners better,and the the labels images that generated by glc-sLDA are similar to the original fastener images.Compared with sLDA model,the proposed algorithm has many obvious advantages.For example,the distinctions between different categories are enhanced and the images of the same category are more similar.Furthermore,the false detection rate is reduced by 55.6% in the fastener image classification.