3.2.2. Cycle Consistency Loss

In actual production, it is easier to obtain a large number of defect-free images, because real defect images are difficult to collect. In addition, there is no difference between a defect image and a defect-free image, except for the defective area. Therefore, the best method for generating defect images is based on defect-free images to assist in defect generation, rather than directly generating defect images [42]. In order to realize the function of defect-free-image-assisted defect image generation, this paper introduced cyclic consistency loss into the network, as shown in Formula (4):

In Formula (4), for the generator C and the defect image G(f) generated by G are taken as inputs to generate the reconstructed pseudo-defect-free image C(G(f)), which is close to the real defect-free image f, where the measure is the L1 norm. Similarly, for generator G, the defect-free image C(i) generated by C is taken as input to reconstruct and generate the pseudo-defect image G(C(i)), which is similar to the real defect image i. The reconstructed defect-free image C(G(F)) is finally similar to the input defect-free image f, and, as a result, the generated defect-free image G(f) maintains its similarity to the input defect-free image f in the defect-free region. By using cyclic consistency loss, DG-GAN can preserve the common features of both defect images and defect-free images.