Deep-learning based segmentation algorithm for defect detection in magnetic particle testing images

Abstract

Magnetic Particle Testing (MPT), also referred to as magnetic particle inspection, is a nondestructive examination (NDE) technique used to detect surface and slightly subsurface flaws in most ferromagnetic materials such as iron, nickel, and cobalt, and some of their alloys. In a bad environment, the procedure is complicated, and automation of MPT is strongly desired. To find defects in the formed magnetic powder pattern, it is required to be highly skilled and automation has been considered difficult. In recent years, many defect detection methods based on deep learning have been proposed, and the effectiveness of deep learning has been shown in the task of automatically detecting various types of defects having different shapes and sizes. In this paper, we describe the development of deep learning based segmentation algorithm for defect detection in MPT images. We have achieved a F2 score of 84.04% by using U-Net as the segmentation model and by utilizing a strong backbone network and an optimal loss function.