Multi-Cue Anomaly Detection and Localization under Data Contamination

TL;DR

This work tackles visual anomaly detection under realistic data conditions where training data are partially contaminated and only a few labeled anomalies are available. It introduces a contamination-robust deviation learning backbone coupled with a tri-signal anomaly scoring scheme and gradient-based localization, enabling reliable detection and interpretable spatial explanations. The approach leverages synthetic pseudo-anomalies, few-shot real anomalies, and adaptive sample weighting to maintain robustness, achieving state-of-the-art performance on MVTec AD and VisA across varying contamination levels, with strong localization and interpretability. The results demonstrate practical impact for industrial inspection by delivering robust, data-efficient, and explainable anomaly detection suitable for deployment in high-stakes settings.

Abstract

Visual anomaly detection in real-world industrial settings faces two major limitations. First, most existing methods are trained on purely normal data or on unlabeled datasets assumed to be predominantly normal, presuming the absence of contamination, an assumption that is rarely satisfied in practice. Second, they assume no access to labeled anomaly samples, limiting the model from learning discriminative characteristics of true anomalies. Therefore, these approaches often struggle to distinguish anomalies from normal instances, resulting in reduced detection and weak localization performance. In real-world applications, where training data are frequently contaminated with anomalies, such methods fail to deliver reliable performance. In this work, we propose a robust anomaly detection framework that integrates limited anomaly supervision into the adaptive deviation learning paradigm. We introduce a composite anomaly score that combines three complementary components: a deviation score capturing statistical irregularity, an entropy-based uncertainty score reflecting predictive inconsistency, and a segmentation-based score highlighting spatial abnormality. This unified scoring mechanism enables accurate detection and supports gradient-based localization, providing intuitive and explainable visual evidence of anomalous regions. Following the few-anomaly paradigm, we incorporate a small set of labeled anomalies during training while simultaneously mitigating the influence of contaminated samples through adaptive instance weighting. Extensive experiments on the MVTec and VisA benchmarks demonstrate that our framework outperforms state-of-the-art baselines and achieves strong detection and localization performance, interpretability, and robustness under various levels of data contamination.

Figures (6)

• Figure 1: Overview of the proposed framework. During training, a segmentation branch, a deviation module, and an uncertainty (entropy) module are learned using synthetic anomaly generation and limited anomaly supervision. At inference, anomaly detection is obtained via tri-signal score fusion, while anomaly localization is produced by a separate gradient-based fusion module that aggregates signal-specific gradients to generate a spatial heatmap.
• Figure 2: Anomaly localization examples showing input images, ground-truth masks, predicted heatmaps, predicted masks, and segmentation outputs for both MVTec AD and VisA.
• Figure 3: Robustness evaluation: Image-level and pixel-level AUROC under varying anomaly contamination rates on the MVTec and VisA datasets.
• Figure 4: Distributions of deviation, uncertainty, segmentation, and fused anomaly scores for nominal and anomalous samples. Each score highlights different aspects of anomalous behavior, illustrating their separability.
• Figure 5: Drop (%) in image and pixel-level AUROC between consecutive contamination levels (5%$\rightarrow$10%, 10%$\rightarrow$15%, and 15%$\rightarrow$20%). The plots summarize sensitivity to contamination on (a)–(b) MVTec and (c)–(d) VisA.