Self-Supervised Railway Surface Defect Detection with Defect Removal Variational Autoencoders

In railway surface defect detection applications, supervised deep learning methods suffer from the problems of insufficient samples and an imbalance between positive negative samples. To overcome these problems, we propose a lightweight two-stage architecture including cropping network (RC-Net) defects removal variational autoencoder (DR-VAE), which requires only normal for training to achieve detection. First, design simple effective RC-Net extract surfaces accurately inspection images. Second, DR-VAE is proposed background reconstruction images detect by self-supervised learning. Specifically, during process, contains random mask module (D-RM) generate signals uses structural similarity index measure (SSIM) as pixel loss. addition, decoder also acts discriminator implement introspective adversarial training. inference stage, reduce error introducing distribution capacity attenuation factor, finally use residuals original reconstructed segmentation defects. The experiments, core parameter exploration comparison with other models, indicate that model can high accuracy.