Robust Distribution Alignment for Industrial Anomaly Detection under Distribution Shift

TL;DR

Industrial anomaly detection often fails under distribution shift due to lighting, sensor drift, or corruptions. This paper proposes RoDA, a memory-bank based framework that performs robust distribution alignment using discretized Sinkhorn distance (robust OT) and target-domain data augmentation to align target patches with source normal prototypes without requiring target labels or decoders. RoDA demonstrates strong improvements over state-of-the-art anomaly detectors and domain adaptation methods on 2D (MVTec, RealIAD) and 3D (MVTec-3D) benchmarks under simulated shifts, including both detection and segmentation tasks. The results highlight the importance of robust OT-based distribution alignment and targeted augmentation for practical industrial deployment with limited target data.

Abstract

Anomaly detection plays a crucial role in quality control for industrial applications. However, ensuring robustness under unseen domain shifts such as lighting variations or sensor drift remains a significant challenge. Existing methods attempt to address domain shifts by training generalizable models but often rely on prior knowledge of target distributions and can hardly generalise to backbones designed for other data modalities. To overcome these limitations, we build upon memory-bank-based anomaly detection methods, optimizing a robust Sinkhorn distance on limited target training data to enhance generalization to unseen target domains. We evaluate the effectiveness on both 2D and 3D anomaly detection benchmarks with simulated distribution shifts. Our proposed method demonstrates superior results compared with state-of-the-art anomaly detection and domain adaptation methods.

Figures (5)

• Figure 1: Upper: Examples of synthesized distribution shifts hendrycks2019benchmarking on MVTec and RealIAD. Lower: Results from top baselines and our method. The Upperbound is trained and tested on the original distribution. GNL cao2023anomaly performs well on "Contrast" shift due to AutoContrast augmentation during training.
• Figure 2: Illustration of pipeline of RoDA. The source domain fitting stage constructs a memory bank of normal training features, which serves as a reference for anomaly detection. In the domain adaptation stage, a limited amount of target domain data is augmented and aligned with the source memory bank through robust optimal transport.
• Figure 3: Illustration of distribution alignment via moment matching, optimal transport and finally our modified robust sinkhorn distance.
• Figure 4: Qualitative results for anomaly segmentation. We present results for PatchCore without adaptation (w/o Adapt), RD4AD, RoDA (Ours) and predictions on original distribution testing sample as Upperbound. RoDA consistently improves anomaly localization compared to the baseline (w/o Adapt), even approaching the Upperbound in some cases.
• Figure 5: (a) Comparison of anomaly sample assignments across different strategies. (b) Evaluation on different severity levels of Gaussian Noise Corruption.