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.
翻译:由于正常和异常数据数量极不平衡,人们越来越注意未经监督的缺陷检测方法。虽然最近广泛研究了以重建为基础的现有方法,但由于各种纹理表面缺陷检测的完善重建模式,由于同质和非常规表面质素,建立各种纹理表面缺陷检测的强大重建模式仍然是一项艰巨的任务。在本文件中,我们提出了一个新的未经监督的基于重建的方法,称为正常参考关注和缺陷特征认知网络(NDP-Net),以准确检查各种文本缺陷。与大多数基于重建的方法不同,我们的NDP-Net首先使用一个编码模块,从表面纹质中提取出多种规模的歧视性特征,随着拟议人工缺陷和新型像素水平缺陷认知损失而增强的缺陷诊断能力。随后,我们提出了一个新的基于参考模块(RBAM),以利用固定参考图像的正常特征来修复缺陷并限制缺陷的重建。随后,修复的特征被注入到一个解码模块中,提取了表层结构结构质变的多级化模块,用于重建正常的测试。最后,一个用于改进SMS的常规测试部分。