Defect Information Synthesis via Latent Mapping Adversarial Networks

This research presents a new image synthesis methodology for automated visual inspection (AVI) in steel manufacturing process. We develop a novel methodology, termed Latent Mapping Adversarial Networks. As the end product of the manufacturing process is directly linked to economic factors, various methods are being utilized to improve the quality of the product. Among them, the defect detection steps carried out in advance are important as it greatly impacts productivity. However, new challenges have emerged for several reasons. First, it requires prior knowledge of the expert to define the defect image and perform detection. To alleviate this problem, various companies have started utilizing AVI to reduce this dependence on domain knowledge. Secondly, defect detection is an arduous task since fewer defect images are available compared to normal images. This underlying problem leads to a classification model that is biased toward the majority class, which degrades the final performance. In this paper, we propose a method to synthesize defect images to solve the above-mentioned problems. Inspired by StyleGAN, we build mapping networks for latent space of the generator. Through this, we can synthesize defect images of various sizes in the manufacturing process. In addition, we experiment to find the most suitable loss function to solve the common problems of Generative Adversarial Networks (GAN). We also optimized the proposed method in terms of convergence and computation speed by estimating the size of optimal latent space. The experimental results using quantitative metrics illustrate the improved performance of the proposed methodology. As a result, it is now possible to solve the quality problem and increase productivity by reducing misclassification in the model through AVI experiments using the generated images