Identification And Analysis Of Tread Surface's Damage Feature Based On Image

With the development of railway industry,as a critical part of the train,the detection and maintenance of wheel treads attract more and more attention.At present,impact loading method belonging to contact measurement and visual-analysis method are the primary method in practical application,while these two methods both have low efficiency in practical application.Machine vision,as an important nondestructive testing method,which could play the role of visual-analysis,has been widely used for its non-contact,fast,high efficiency and stability.This paper takes the wheel treads as the measurement objects to study nondestructive testing of the treads' surface condition based on machine vision.Based on current images-texture-feature-extraction methods,combining statistical methods and the spectral methods,a method is proposed which includes selecting features based on SVM firstly,and then classifying and recognizing features based on K-means.What matters in this method is the selection of texture feature vectors.Among current machine learning methods,firstly the characteristics of supervised learning and unsupervised learning are analyzed separately,and SVM is selected which belongs to supervised learning,to perform feature selection Next,according to the texture characteristics of wheel tread,feature extraction methods such as GLCM,GGCM,Gabor and Haar wavelet transform are studied,and the application effects of various parameters are also analyzed.At the same times,texture features which have same physical meaning or have large redundancy after PCA,will be screened by difference method or feature selection algorithm based on SVM.Lastly,a multidimensional feature vector that can reflect the local information of image gray scale gradient cue and the multi-scale frequency domain information can be gotten.This paper also designs many other algorithms to merge defects,to make morphological determination and to screen preliminarily.Combining above algorithms with K-means clustering algorithm,finally,a nondestructive testing algorithm of tread images' surface condition based on K-means and texture feature vectors is proposed.Based on this nondestructive testing algorithm,after analyzing the detection results of actual wheel treads,it is proved that the detection results of wheel treads have high-accuracy and computational-speed.Which means this algorithm is timely-and-accurately.