Absolute-Unified Multi-Class Anomaly Detection via Class-Agnostic Distribution Alignment

TL;DR

This work tackles absolute-unified multi-class unsupervised anomaly detection by removing dependence on class labels during training and inference. It introduces Class-Agnostic Distribution Alignment (CADA), a regression-based module that predicts per-image normal-score distribution statistics $(\hat{u},\hat{\gamma})$ and normalizes anomaly maps accordingly, enabling unified detection across all classes. CADA can be plugged into existing UAD methods with modest overhead and is shown to yield substantial performance gains on MVTec AD and VisA, outperforming previous state-of-the-art by large margins (e.g., I-AUROC rising from 91.4% to 98.6% on MVTec AD). The approach addresses inter-class distribution mismatch without requiring class information, offering a practical path to scalable, class-agnostic anomaly detection in industrial and medical contexts.

Abstract

Conventional unsupervised anomaly detection (UAD) methods build separate models for each object category. Recent studies have proposed to train a unified model for multiple classes, namely model-unified UAD. However, such methods still implement the unified model separately on each class during inference with respective anomaly decision thresholds, which hinders their application when the image categories are entirely unavailable. In this work, we present a simple yet powerful method to address multi-class anomaly detection without any class information, namely \textit{absolute-unified} UAD. We target the crux of prior works in this challenging setting: different objects have mismatched anomaly score distributions. We propose Class-Agnostic Distribution Alignment (CADA) to align the mismatched score distribution of each implicit class without knowing class information, which enables unified anomaly detection for all classes and samples. The essence of CADA is to predict each class's score distribution of normal samples given any image, normal or anomalous, of this class. As a general component, CADA can activate the potential of nearly all UAD methods under absolute-unified setting. Our approach is extensively evaluated under the proposed setting on two popular UAD benchmark datasets, MVTec AD and VisA, where we exceed previous state-of-the-art by a large margin.

Figures (6)

• Figure 1: Task settings of (a) class-separated (conventional) UAD, (b) model-unified multi-class UAD you2022unified, and (c) absolute-unified multi-class UAD (proposed).
• Figure 2: Application scenarios of absolute-unified multi-class UAD. (a) Components of a product are mixed on the assembly line. (b) A product has various types on a flexible manufacturing line. (c) Mixed inspection views for different parts of a large product.
• Figure 3: A toy example of why methods fail in absolute-unified UAD. First row: Normal and anomalous samples cannot be separated when three classes are mixed. Second row: For each class, the anomaly scores are normalized by the mean (${u}_{c}$) and maximum (${\gamma}_{c}$) of the normal samples; hence, the unified distribution is separable. Under absolute-unified setting, ${u}_{c}$ and ${\gamma}_{c}$ cannot be directed computed as images are "anonymous". CADA can estimate and align the distribution without knowing image classes.
• Figure 4: Anomaly score alignment for absolute-unified UAD. (a) Class-aware. Class labels are provided in training and test phase. (b) Class-aware. Class labels are provided only in training phase. (c) Class-agnostic. Class label not provided (CADA, Ours).
• Figure 5: Image-level anomaly score distribution on MVTec AD. (a) RevDist. (b) RevDist+CADA. The distributions of normal samples are more consistent across different classes with the proposed CADA.