Attention Guided Deep Neural Network for Inspection of Surface Defects on Steel Products

Recent advancements in deep learning, particularly in semantic segmentation, have achieved notable success in various industrial applications. However, the characteristics of images vary between applications, presenting distinct challenges. In the case of steel images, environmental factors during acquisition significantly affect their appearance, and the random occurrence of defects complicates their generalization, making defect segmentation for surface inspection particularly challenging. This paper introduces TAG-Net, a novel attention-based semantic segmentation network aimed at improving the distinction between background and defects in challenging input images. TAG-Net estimates three attention maps for the background, defects, and their boundaries, with boundary detection included as an auxiliary task to enhance the guidance of the attention maps. Experiments on the NEU-Seg dataset demonstrate that our proposed method significantly outperforms traditional baseline approaches for general images and recent approaches for steel images, yielding superior segmentation performance.