LAUSR

In the context of industrial product quality control, visual defect detection plays a crucial role, emphasizing the importance of developing a high-performing defect detection model. Thus, it is imperative to propose an exceptional defect detection model that can ensure the reliable and accurate detection of defects. In this article, we introduce a novel unsupervised reconstruction-based method, called the normal reference attention and defective feature perception network (NDP-Net), which accurately detects various textured defects. Unlike most reconstruction-based methods, our NDP-Net first employs an encoding module that extracts multiscale discriminative features of the surface textures, which is augmented with the discriminative defect ability by proposed artificial defects and a novel pixel-level defect perception loss. Subsequently, a novel reference-based attention module (RBAM) is proposed to leverage the normal reference features to repair the defective ones and restrain the reconstruction of the defects. Finally, a novel multiscale defect segmentation module (MSDSM) is introduced for precise defect detection and segmentation. Exhaustive experimental results obtained from benchmark datasets, including public MVTec AD and DAGM, showcase the exceptional performance and efficiency of the proposed method. Furthermore, we successfully implemented the proposed method in practical industrial settings, demonstrating its efficacy in real-world applications.