Defect image sample Generation for Automatic rail defect Recognition
Abstract - Railway defects can lead to enormous economic and human loss. Among all the defects, surface defects are most common and prominent type, and there have been many attempts to detect them with optical-based non-destructive testing (NDT) methods. However, there are many sources of errors that can jeopardize the NDT operation, among which human errors are the most unpredictable and frequent. The application of artificial intelligence (AI) to NDT interpretation has the potential to minimize human intervention, but the lack of sufficient images of the railways with various type of defects is the major obstacles to train the AI models. To address the issue of data scarcity, RailGAN model was proposed, which can improve the Cycle GANmodel by introducing an additional pre-sampling stage for the railway tracks in the algorithm. Two pre-sampling techniques were tested for RailGAN model: image-filtration and U-Net. By applying both techniques to 20 real-time railway images, U-Net could produce more consistent results of performing image segmentation on all the images while being less affected by pixel intensity values. Adoption of U-Net for the RailGAN model and applying both the RailGAN and the original CycleGAN models to the same real-time railway images showed that the original CycleGAN model tends to generate defects in the irrelevant background, whereas the RailGAN adds these synthetic defect patterns only on the railway surface. These images resemble real cracks on the railways and can be used for the training of neural network-based defect identification algorithms
Keywords - Railway Defects, Nondestructive Testing, Image Augmentation, Machine Learning.