Reserch On Structrual Semantic Learning Method Of ST-LDA Oriented To Features In Railway Fastener Images

A Railway fastener is a kind of component used to hold the track on the sleepers. To detect railway fasteners via computer vision techniques, which is growing increasingly hot, instead of manual inspection, has become a promising measure to achieve intellectualization in the department of railway maintenance. Classifiers estimate the state of a fastener directly based on low level features extracted from images is the present common fastener detection process based on computer vision. Because of the weak robustness of low level features and the boundary between the categories of valid and invalid fastener images is difficult to fit in the classifier, present processes could hardly achieve ideal detection results. In this paper, during the interval of features and classifier, topic distributions were obtained by means of semantic learning of image features. Then these distributions were input into the classifiers to decide the category of a fastener image. Traditional semantic methods initiate the learning processes by encoding the features into words. Unfortunately, the strategy of this encoding ignores the structural information in images. In this paper, a semantic approach, which is able to provide cues of fastener structural conditions, is proposed to learn features. Specifically, main studies of this paper are as follows:(1) Aimed at the problem that the traditional method of LBP(local binary pattern) could not robustly describe the structure of a fastener in the condition of strong illumination varying, a binary LBP method,which is a modified LBP, was proposed to extracted the structural image of fasteners.The main characteristic of the binary method is that a certain orient along which the fastener shape could be depicted most perfectly, was selected among the 8 orients of traditional LBP, and then original images were encoding to structure images along the certain orient. Theoretical analysis manifested that probability of illumination robustness was enhanced by the proposed method. The rate of miss was 7.7% and the false alarm was 11.3 during the experiment that was conducted based on structural images. The experiment result demonstrated that the structural images could robustly describe the structure and the shape of fasteners.(2) Aimed at the ignorance of structural information in traditional LDA, Based on the aforementioned structural images, a variable of structure was configured in LDA, and a structural semantic learning method of St_LDA(structure latent dirichlet allocation), which is a modified LDA method, was proposed in order to extract topic distributions that could indicate the stage of fastener structure from features. The topic distributions were used to be the inputs of classifiers. Firstly, features were encoded in terms of structure variables according to the location of the features in structural images. Therefore, the traditional unidimensional word histograms were expanded to two-dimensional word-structure encodings. Then, a St_LDA generative mechanism was built for the two-dimensional encodings, and the formulas used to learning features were derived according to the mechanism and the topic distributions were estimated through the formulas. Experiments results showed that, compared with classical LDA, the inter-class distance between valid and invalid fasteners increased by 5%～3% under the condition of St_LDA; When the St_LDA learning was combined with the present fastener detection processes, the rates of miss decreased by approximately 76% and the false alarm decreased by 67%-72%.