Abstract: This paper tackles the task of defect segmentation by exploiting sufficient normal (defect-free) training images and limited annotated anomalous images. We propose a class activation map ...