Self-Supervised Structure Learning for Crack Detection Based on Cycle-Consistent Generative Adversarial Networks

Deep learning is a state-of-the-art approach to pixel-level crack detection. However, it relies on a large number of source-target image pairs for the training, which is very expensive. This paper proposes a self-supervised structure learning network which can be trained without using paired data, even without using ground truths (GTs); this is achieved by training an additional reverse network to translate the output back to the input simultaneously. First, a labor-free structure library is prepared and set as the target domain for structure learning. Then a dual network is built with two generative adversarial networks (GANs); one is trained to translate a crack image patch (X) to a structural patch (Y), and the other is trained to translate Y back to X, simultaneously. The experiments demonstrated that with such settings, the network can be trained to translate a crack image to the GT-like image with a similar structure pattern, and it can be used for crack detection. The proposed approach was validated on four crack data sets and achieved comparable performance to that of state-of-the-art supervised approaches.