loading page

Distillated Spatial-Attention Reconstruction Module for Surface Defect Detection of Industrial Products
  • +3
  • Hao WAng ,
  • Xiyu He ,
  • Peiyuan Zhu ,
  • Jiaxing Zeng ,
  • Xiaohao Wang ,
  • Xian Qian
Peiyuan Zhu
Author Profile
Jiaxing Zeng
Author Profile
Xiaohao Wang
Author Profile
Xian Qian
Tsinghua-Shenzhen International Graduate School

Corresponding Author:[email protected]

Author Profile

Abstract

To fully address the problem for surface defect detection in real product line, we proposed an abnormal detection method based on Distillated Spatial-Attention Reconstruction (DSAR) module and apply it to the industrial magnetic tile surface quality inspection line. The proposed structure is a reconstruction-based algorithm, which applies distillation module and spatial-attention module to reduce the generalization of reconstruction network. Compared with existing methods, DSAR achieves the best results on our magnetic tile dataset and solved the critical long crack defect that the AUROC is improved from 88.6 to 98.6. On the public dataset MVTec, DSAR gets the best results of the reconstruction-based method and the third of all methods. We experimentally verify the role of distillation module and spatial attention module. We believe that our method also can be applied to the inspection of many types of industrial products in real product lines.