.png){kind=icon}
Return
Reconstruction error-aware collaborative memory network for unsupervised anomaly detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.113998.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The powerful image generation capabilities of generative adversarial network (GAN) can be leveraged for input reconstruction in unsupervised anomaly detection. GAN-based unsupervised anomaly detection is capable of effectively reconstructing input samples and the quality of these reconstructed samples has a non-negligible impact on the overall anomaly detection performance. However, existing methods often suffer from residual anomaly information within the latent feature space, preventing the normalized reconstruction of anomalous regions. To address this limitation, we propose a collaborative memory module, which transforms image features into multiple memory features that record distinct patterns through parallel memory units, thereby acquiring high-quality normal feature representations. Furthermore, existing methods often exhibit insufficient information interaction in the encoding process of intermediate layers between the input and reconstructed samples, leading to information loss and poor image detail. To mitigate this issue, we propose a reconstruction error guided network, which collects reconstruction error information from both encoders and employs a grouped reconstruction error attention mechanism to guide the reconstructed image encoding process, thus preserving more input image detail. Extensive experiments on multi-normal class anomaly detection tasks and industrial defect detection tasks demonstrate the promising performance of our proposed reconstruction error-aware collaborative memory network.
Keywords:
Anomaly detection
Collaborative memory
Reconstruction error
Information interaction
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
No organization information available
Cited Papers
No cited papers available