DMAD: Dual Memory Bank for Real-World Anomaly Detection

TL;DR

The results show that DMAD surpasses current state-of-the-art methods, highlighting DMAD's capability in handling the complexities of real-world anomaly detection scenarios.

Abstract

Training a unified model is considered to be more suitable for practical industrial anomaly detection scenarios due to its generalization ability and storage efficiency. However, this multi-class setting, which exclusively uses normal data, overlooks the few but important accessible annotated anomalies in the real world. To address the challenge of real-world anomaly detection, we propose a new framework named Dual Memory bank enhanced representation learning for Anomaly Detection (DMAD). This framework handles both unsupervised and semi-supervised scenarios in a unified (multi-class) setting. DMAD employs a dual memory bank to calculate feature distance and feature attention between normal and abnormal patterns, thereby encapsulating knowledge about normal and abnormal instances. This knowledge is then used to construct an enhanced representation for anomaly score learning. We evaluated DMAD on the MVTec-AD and VisA datasets. The results show that DMAD surpasses current state-of-the-art methods, highlighting DMAD's capability in handling the complexities of real-world anomaly detection scenarios.

Figures (6)

• Figure 1: (a). Comparison of Anomaly Detection's Settings. The current research trend is moving from one-class to unified (multi-class) and from unsupervised to semi-supervised, which is more practical. We propose a unified with a few annotated anomalies setting, filling the gap in this field. (b). Defects in various objects often exhibit visual similarities. Therefore, defect data from one object can aid the model in detecting similar defects in other objects within a unified semi-supervised setting.
• Figure 2: Overview of DMAD. DMAD is a unified framework that accommodates both unsupervised and semi-supervised scenarios. During training, normal images and seen annotated anomalies are input into the Feature Extractor to obtain patched features. For the anomalies, a $Filter$ is utilized to isolate the anomalous parts. Subsequently, a Dual Memory Bank Knowledge Enhancement is employed to obtain enhanced representations of both normal and abnormal. A more detailed illustration is depicted in \ref{['fig:framework_detail']}. A pseudo abnormal representation is generated by utilized a feature augmentation liu2023simplenet. Ultimately, these normal representations, abnormal representations, and pseudo abnormal representations are used to train a MLP.
• Figure 3: An intuitive illustration of the Dual Memory Bank's construction is depicted on the left. A schematic diagram of the Knowledge Enhancement based on the Dual Memory Bank is shown on the right. DMAD uses the Dual Memory Bank to create an enhanced representation encompassing normal and abnormal knowledge.
• Figure 4: Qualitative comparison of our method with UniAD you2022unified, BGAD yao2023explicit and SimpleNet liu2023simplenet on MVTec-AD (Left) and VisA (Right) dataset.
• Figure 5: I-AUROC comparison for varying numbers of anomalies on MVTec.