Industrial Surface Anomaly Detection by Combining Grayscale Edge Features and Normal Memory Features of Images for Reconstruction

• Yao Yang,
• Xiaoling Chen,
• Linna Zhang,
• Yansen Huang,
• Yi Cen,
• Yigang Cen,

Vol. 19, No. 3, March 31, 2025
10.3837/tiis.2025.03.010, Download Paper (Free):

• deep neural network
• Memory module
• feature fusion
• Anomaly simulation

Abstract

Reconstruction-based anomaly detection methods compare reconstructed image and input image and utilize the differences to determine the anomaly. But deep neural networks are prone to overgeneralization, and thus may result in anomalous regions being reconstructed as well as normal regions, which will decrease the anomaly detection performance of the model. An industrial surface detection reconstruction network (GAMRec) that fuses grayscale edge features and its most similar normal memory features of an image for reconstruction, which can well alleviate the overgeneralization problem. In specific, this is achieved with a UNet-type autoencoder with an attention module, an anomaly simulation module and a memory module that does not need to be updated. The grayscale edge features contain rich low-frequency information well, while the normal memory features contain the rich detail information. The fusion of these two different features can be utilized to reconstruct more realistic and comprehensive images. The proposed task can effectively prevent the reconstruction of abnormal regions. Experiments on MVTec AD and BTAD public datasets and our sewer dataset Sewer-AD validate that GAMRec achieves superior performance in detecting and localizing anomalies for industrial product surface anomaly detection.

Statistics

Cite this article