Normal Image Generation-Based Defect Detection by Generative Adversarial Network with Chaotic Random Images

Hiroki Kobayashi, Ryo Miyoshi, Manabu Hashimoto

Abstract: We propose a defect detection method called ChaosGAN (Generative Adversarial Network with Chaotic Random Images) for image generation that can output a normal image with high reconstruction performance regardless of whether the input image is normal or anomaly. A defect detection method based on image generation should be able to (i) convert from a normal image to the same normal image as input and (ii) convert from an anomaly image to normal image with the defective parts removed. ChaosGAN is a combination of Skip-GANomaly, which performs well at identity mapping of a normal image, and AnoGAN, which reconstructs a normal image by regarding a random image as an input latent space. We conducted an experiment to evaluate ChaosGAN using the area under the curve of receiver operating characteristic (AUROC). The AUROC was 0.76 with ChaosGAN (AnoGAN was 0.49, and Skip-GANomaly was 0.67), indicating that it performs better than other defect detection methods.