CNN-DFT Based Approach Applied to Image Inspection of Railcar Component: A Comparison with Machine Learning Methods

Abstract

The railcar component inspection is one of the most critical tasks in railway maintenance. The use of image processing, coupled with machine learning has emerged as a solution for replacing current standard methodologies. The spectral analysis gives the frequency representation of a signal and has been largely used in signal processing tasks. In this sense, this work proposes the evaluation of the use of the discrete Fourier transform (DFT) in addition to the spatial representation image of railcar component for an automatic detector of defective parts performed by convolutional neural network (CNN) classification. The most appropriate combination of images of the spatial and frequency domains is compared to the histogram of oriented gradients (HOG) feature descriptor linked to the multilayer perceptron (MLP) and support vector machine (SVM) classification, where data augmentation is investigated to improve the classification performed by all approaches. A search is made for the parameters that best fit the MLP and SVM models for comparison with the proposed approach. The results are given in measure of accuracy in addition to accuracy boxplot, and it showed encouraging results in the combination of spatial image and DFT magnitude combined with data augmentation as CNN inputs, reaching an accuracy of 96.04% and demonstrating statistically to have a significant difference between the comparative methods.