Railway Fastener Detection Based On Convolution Neural Network

In the process of Railway development and operation, railway infrastructure testing has always been the focus of railway safety, which is very important part of the fastener state detection. In recent years, with the rapid development of computer technology and image processing technology. Based on Computer Vision missing fasteners automatic detection system has become an important measure to maintain the environmental safety of railway, a hot topic in the field of visual inspection.In previous group proposes a fastener detection method based on online learning. The method through online learning strategy, dynamically generated template library, so no training, it can be easily extended to different lines. Currently, deployed in the implementation of Railway Science Research Institute of the track inspection, it has accumulated a large amount of data, and achieved good detection results. However, this method also has some shortcomings:First, this method requires periodic maintenance fasteners static template library, and secondly, because there is no training, in relative terms, a high error detection rate.In order to further improve the detection efficiency, while leveraging a large number of pre-acquisition and analysis of data obtained, this paper presents an abnormal state detection based rail fastener convolution neural network approach. With the different of previous manual HOG feature, deep learning is a feature extraction method based learning, training through the early accumulation of large amounts of data, it can effectively improve the detection accuracy. And for some special fastener abnormal state, such as:small changes fasteners, loose and elastic fastener strip breakage, etc., can also be done to detect different line rail fastener through re-learning and training.The main contributions of this thesis are as follows:1. Preparation of data sets:This paper has been based on the previous research group online learning adaptive fastener testing procedures, completed the acquisition of rail fastener data, the fastener data sets are divided into two categories. One classification is normal, another classification is abnormal. Three lines collected data rail fasteners, which are set through the Jitong line, the Huning and Huhang high-wire line. Jitong line total 47281 images fastener, wherein the fastener 24670 normal, abnormal fastener 22611. Huning line total 27869 images fastener, wherein the fastener 13946 normal, abnormal fastener 13923. Huhang high-wire a total of 35428 images fastener, wherein the fastener 18456 normal, abnormal fastener 16972.2. From the image classification perspective, an abnormal state is detected fastener normal and abnormal state of binary classification. To this end, we propose a method for detecting fastener AlexNet Network Model, which is a feature based method of learning by progressively extracted by the low-level features, can be more abstract high-level categories of characteristics, to overcome the manual HOG features and existing fastener template library complex fastener weak change adaptation issues, to changes in the light and caused stains at Jitong line of fasteners can be achieved very good result.3. From the image verification perspective, fasteners abnormality detection problem can be seen as validation problems be detected with the known classification fastener fasteners, it is the same as normal, except the abnormality occurs. To this end, we propose a method Siamese Network Model based on the input of a pair of images, both are normal or abnormal state of the image to be similar images were normal and the abnormal state of the image is dissimilar images. By building on similar and dissimilar to the training set, the fastener can be characterized by a more discriminative training and learning. This method in the constructed data set of Huning and Huhang line was validated detection method based on the results better than the fastener Classification.