Unsupervised Automatic Defect Inspection Based on Image Matching and Local One-Class Classification

Chengkan Lv, Zhengtao Zhang, Fei Shen, Feng Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 4435-4444

Abstract


In this paper, an unsupervised defect inspection method based on anomaly detection is proposed to inspect various kinds of surface defects in the field of industrial production. This method consists of two modules: (i) An image matching module is utilized to align the input image with a pre-specified template image. Specifically, all objects to be detected will be adjusted to the same position and angle. The aligned images can reduce the difficulty of the training stage, facilitating the subsequent feature extraction and anomaly localization. (ii) After the image matching procedure, an anomaly localization module is trained to learn a mapping that concentrates normal samples in feature space. In particular, each local image region is assigned a feature center by adopting a feature map as the mapping target. Therefore, the compactness of the features extracted from the same region can be improved, which is beneficial to detect potential anomalous targets. Moreover, various artificial defective images are synthesized during the training stage to further improve the discriminatory ability of the anomaly localization module. A series of experiments are conducted on MAD dataset and the industrial production line. The experimental results verify the efficiency and versatility of the proposed method.

Related Material


[pdf]
[bibtex]
@InProceedings{Lv_2023_CVPR, author = {Lv, Chengkan and Zhang, Zhengtao and Shen, Fei and Zhang, Feng}, title = {Unsupervised Automatic Defect Inspection Based on Image Matching and Local One-Class Classification}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2023}, pages = {4435-4444} }