Filter or Compensate: Towards Invariant Representation from Distribution Shift for Anomaly Detection

TL;DR

FiCo firstly compensates the distribution-specific information to reduce the misalignment between the teacher and student network via the Distribution-Specific Compensation module, and secondly filters all abnormal information to capture distribution-invariant normality with the Distribution-Invariant Filter (DiIFi) module.

Abstract

Recent Anomaly Detection (AD) methods have achieved great success with In-Distribution (ID) data. However, real-world data often exhibits distribution shift, causing huge performance decay on traditional AD methods. From this perspective, few previous work has explored AD with distribution shift, and the distribution-invariant normality learning has been proposed based on the Reverse Distillation (RD) framework. However, we observe the misalignment issue between the teacher and the student network that causes detection failure, thereby propose FiCo, Filter or Compensate, to address the distribution shift issue in AD. FiCo firstly compensates the distribution-specific information to reduce the misalignment between the teacher and student network via the Distribution-Specific Compensation (DiSCo) module, and secondly filters all abnormal information to capture distribution-invariant normality with the Distribution-Invariant Filter (DiIFi) module. Extensive experiments on three different AD benchmarks demonstrate the effectiveness of FiCo, which outperforms all existing state-of-the-art (SOTA) methods, and even achieves better results on the ID scenario compared with RD-based methods. Our code is available at https://github.com/znchen666/FiCo.

Figures (5)

• Figure 1: Anomaly map from different scenarios of SOTA AD methods bae2023pniliu2023simplenettien2023revisitingcao2023anomaly on the MVTec benchmark bergmann2019mvtec. The image of each row represents a different scenario, including ID and four OOD (brightness, contrast, defocus blur and gaussian noise) scenarios.
• Figure 2: The overall architecture of our method FiCo, including detailed structure of DiSCo module and DiIFi module with designed losses. DiSCo module aims to compensate for the distribution-specific information by $\mathcal{L}_{Co}$ to prevent misalignment. DiIFi module attempts to filter abnormal patterns to obtain invariant normality with $\mathcal{L}_{Fi}$, including $\mathcal{L}_{lowf},\mathcal{L}_{mse},\mathcal{L}_{nor}$. $\mathcal{L}_{abs}$ indicates the consistency loss between original and augmented representation at OCBE module.
• Figure 3: Experimental results on hyper-parameters on the MVTec benchmark.
• Figure 4: Anomaly map of $f^{D_k}$ and $f_F^{D_k}$ on the MVTec benchmark. Each row represents a different scenario, including ID, defocus blur and gaussian noise. For each scenario, two examples are shown with the original image, the groundtruth label, anomaly map from $f^{D_k}$ and from $f_F^{D_k}$.
• Figure 5: Anomaly scores of FiCo and GNL cao2023anomaly on ID, gaussian noise scenario from the MVTec benchmark. Each scenario consists of two examples from 'zipper', 'toothbrush', and 'pill', 'metal nut'. Color blue and red indicates distribution on normal samples and anomalous samples, respectively.