UNSUPERVISED IMAGE SEGMENTATION FOR RUST DEFECT DETECTION BY INVARIANT INFORMATION CLUSTERING

Abstract

Metal product corrosion is a concern for various products as the appearance of metal products is compromised when red rust or other corrosion occurs, resulting in a decline in product value. In addition, embrittlement, depletion, and destruction of metal materials can result in a loss of strength, causing social effects. In some cases, this can result in serious accidents involving human lives, attracting significant social attention. Protective coatings, such as the widely used zinc plating, are frequently applied to metal products to protect the metal base. When inspecting the corrosion of such metal products, it is desirable to be able to measure not only the areas containing red rust but also those with white rust while monitoring the progress of corrosion simultaneously. However, most corrosion inspections are performed visually and subject to the inspector’s experience. In this study, we propose an unsupervised rust image segmentation method. Inspired by invariant information clustering, we optimize the parameters of a multi-layer convolutional neural network to maximize the mutual information content and achieve rust area segmentation without using annotation data. We investigated the appropriate number of clusters through corrosion experiments on zinc-coated metal plates and realized the image segmentation of red and white rust.