Unsupervised anomaly detection with memory bank and contrastive learning

Using contrastive learning to optimize memory bank based storage features is an important issue in the field of image anomaly detection. Current memory-based anomaly detection methods, which are commonly based on pre-trained feature extractors, frequently encounter difficulties attributable to the suboptimal quality of features. Additionally, many patch-based methods primarily focus on localized feature differences, overlooking the broader contextual information crucial for robust anomaly detection. To address these issues, we propose a novel framework, named MemConNet, which combines contrastive learning to optimize memory features. Specifically, we propose to employ a pre-trained feature extraction backbone and apply Channel-Weighted Feature Recalibration (CWFR) and Multi-Scale Recalibration Pyramid (MSRP) to capture fine-grained details and broader contextual information, enhancing the channel features and ensuring that key features are emphasized. Secondly, we propose to leverage contrastive learning to refine feature distinctions, while utilizing a memory bank to store and compare normal feature representations, achieving higher precision in anomaly detection. Extensive experiments show that our MemConNet achieves 99.3% Image-level AUROC on the MVTec AD dataset and 97.1% Image-level AUROC on the ViSA dataset, which are 0.2% and 1.8% higher than the comparison methods, respectively, demonstrating excellent detection results.