MÆIDM: multi-scale anomaly embedding inpainting and discrimination for surface anomaly detection

The detection of anomalous structures in natural image data plays a crucial role in numerous tasks in the field of computer vision. Methods based on image reconstruction or inpainting are trained on images with no anomalies or artificial anomalies; anomaly detection and localization are achieved by...

Full description

Saved in:

Bibliographic Details
Published in:	Machine vision and applications Vol. 34; no. 4; p. 66
Main Authors:	Sheng, Siyu, Jing, Junfeng, Jiao, Xintian, Wang, Yafei, Dong, Zhenyu
Format:	Journal Article
Language:	English
Published:	Berlin/Heidelberg Springer Berlin Heidelberg 01-07-2023 Springer Nature B.V
Subjects:	Accuracy Algorithms Anomalies Coders Communications Engineering Computer Science Computer vision Datasets Defects Embedding Image Processing and Computer Vision Image reconstruction Image segmentation Industrial production Localization Methods Networks Neural networks Original Paper Pattern Recognition Performance enhancement Pixels Surface defects Vision systems Computer vision Image reconstruction and inpainting Surface anomaly detection Multi-scale feature enrichment
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	The detection of anomalous structures in natural image data plays a crucial role in numerous tasks in the field of computer vision. Methods based on image reconstruction or inpainting are trained on images with no anomalies or artificial anomalies; anomaly detection and localization are achieved by computing the difference between the input image and the reconstructed image. DRÆM trains two sub-networks to reduce over-fitting of synthetic appearances. This method uses an encoder–decoder and an U-Net-like network to detect and locate anomalies. In order to further improve the performance of the model in accurate inpainting of abnormal images and pixel-level segmentation, we propose a multi-scale anomaly embedding inpainting and discrimination model (MÆIDM). The proposed method introduces a trainable multi-scale feature enrichment module (MFEM) in reconstructive sub-network for image inpainting and an attention discriminative sub-network for defect segmentation. In addition, the Gaussian filtering is used to smooth the anomaly score map. Extensive experiments show that our method achieves excellent performance on the anomaly detection dataset MVTec and two unpublished fabric datasets with AUC scores of 98.5% and 98.1% at the image level and pixel level, respectively. Meanwhile, our model further achieves better detection performance on the supervised DAGM surface defect detection dataset, which proves the effectiveness of the method.
ISSN:	0932-8092 1432-1769
DOI:	10.1007/s00138-023-01425-y