Normal Reference Attention and Defective Feature Perception Network for Surface Defect Detection
Visual anomaly detection plays a significant role in the development of industrial automatic product quality inspection. As a result of the utmost imbalance in the amount of normal and abnormal data, growing attention has been given to unsupervised methods for defect detection. Although existing reconstruction-based methods have been widely studied recently, establishing a robust reconstruction model for various textured surface defect detection remains a challenging task due to homogeneous and nonregular surface textures. In this paper, we propose a novel unsupervised reconstruction-based method called the normal reference attention and defective feature perception network (NDP-Net) to accurately inspect a variety of textured defects. Unlike most reconstruction-based methods, our NDP-Net first employs an encoding module that extracts multi scale discriminative features of the surface textures, which is augmented with the defect discriminative ability by the proposed artificial defects and the novel pixel-level defect perception loss. Subsequently, a novel reference-based attention module (RBAM) is proposed to leverage the normal features of the fixed reference image to repair the defective features and restrain the reconstruction of the defects. Next, the repaired features are fed into a decoding module to reconstruct the normal textured background. Finally, the novel multi scale defect segmentation module (MSDSM) is introduced for precise defect detection and segmentation. In addition, a two-stage training strategy is utilized to enhance the inspection performance.