Research On The Inspection Algorithm Of Railway Fastener Based On Bag Of Words Model

Railway fasteners are the important foundational parts in connecting rails and sleepers.Traditional fastener testing relies on manual to detect the flaws of railway tracks,which is lack of accuracy and real-time data capability.In recent years,the rapid development of computer vision technology has provided perfect technology for the automatic detection of railway fasteners.Bag-Of-Words method is an extensively used image representation method.The method firstly clusters the low-level features of the image to generate visual lexicons.And each clustering center represents a visual word.Then each low-level feature of the image is mapped to the visual word with the shortest Euclidean distance from the image to the visual word to produce a word frequency matrix which is used to represent the image content.Lastly,the word frequency matrix is combined with machine learning methods to perform image classification.This article is targeted to improve the mapping process of low-level features to visual words in Bag-Of-Words method.The main research work is as follows:(1)Aiming at the problem of large quantization error in traditional Bag of words(BOW)model when performing the low-level feature coding,a model of railway fastener detection based on homoionym-assignment is proposed.Firstly,the Latent Dirichlet Allocation(LDA)is used to excavate the latent topic distribution induced by the visual words.Secondly,the relative entropy is introduced to measure semantic distance between visual words,so as to obtain semantically related words.And then,the soft-assignment is adopted to realize the mapping between low-level features and some homoionym.Finally,the Support Vector Machine(SVM)is applied to classify a new image.The experiment on four types of fasteners shows that the proposed model can improve the accuracy of the fastener classification effectively.(2)The traditional Bag-of-Words(BOW)mod el considers nothing about contextual semantic information in the spatial field.Focusing on this defect,a fastener inspection model based on contextual semantic information was proposed.Firstly,based on the traditional BOW model,the Gibbs Random Field model was introduced to combine the feature appearance similarity and contextual semantic information in order to acquire more accurate visual words.And then the Latent Dirichlet Allocation(LDA)was used to learn the topic distribution.At last,the Support Vector Machine(SVM)was applied to classify a new image.Experimental results on four kinds of fastener images show that the proposed algorithm can inspect fastener states more precisely.