BASIC STUDY ON RUST IMAGE GENERATION OF WEATHERING STEEL BY GENERATIVE ADVERSARIAL NETWORKS

Abstract

The general investigation method for weathering steel is a visual five-step evaluation of the rust layer. However, since the visual evaluation of the rust layer may differ depending on the inspectors, the simple and accurate evaluation method is required. From this, the rust evaluation method using deep learning has been proposed, but there is a problem that annotated rust images are less. Therefore, in this study, the possibility of rust image generation by GAN was examined by comparing the rust images used as training data with the generated rust images by GAN. The results showed that the rust image generated from the dataset consisting of rust images more than 4 images with the score 3 is judged to have the score 3 by an expert rust evaluation, and have a certain diversity.