Efect classification of railway fasteners using image preprocessing and a lightweight convolutional neural network

Abstract

Railway fasteners are used to securely fix rails to sleeper blocks. Partial wear or complete loss of these\rcomponents can lead to serious accidents and cause train derailments. To ensure the safety of railway transportation,\rcomputer vision and pattern recognition-based methods are increasingly used to inspect railway infrastructure. In\rparticular, it has become an important task to detect defects in railway tracks. This is challenging since rail track images\rare acquired using a measuring train in varying environmental conditions, at different times of day and in poor lighting\rconditions, and the resulting images often have low contrast. In this study, a new method is proposed for the classification\rof defects on rail track fasteners. The proposed approach uses image enhancement to first filter the rail images and obtain\ra high contrast image. Then, the rail track and sleeper positions are determined from the high contrast image. The\rlocation of the fastener is determined by applying the line local binary pattern method and the defects of the fastener\rare classified using an improved lightweight convolutional neural network (LCNN) model. Features are extracted from\rtwo fully connected layers of the developed LCNN model and the feature vector is constructed by concatenating these\rlayers. The concatenated features are processed using a number of machine learning methods and the optimum classifier\ris chosen. Experimental results show that Cubic SVM gives the best results with a detection accuracy rate of 99.7%.