ShortcutBreaker: Low-Rank Noisy Bottleneck with Global Perturbation Attention for Multi-Class Unsupervised Anomaly Detection

TL;DR

MUAD suffers from identity shortcuts that reduce anomaly discrimination when learning a unified model. ShortcutBreaker addresses this with a low-rank noisy bottleneck (LRNB) that prevents trivial identity reproduction and a global perturbation attention (GPA) that disrupts shortcut pathways in the decoder, implemented on a frozen DINO-pretrained ViT encoder with a learnable Transformer decoder. Evaluations on MVTec-AD, ViSA, Real-IAD, and Universal Medical show consistent state-of-the-art image- and pixel-level detection across industrial and medical domains, with LRNB contributing the most and GPA further enhancing robustness. The work offers a practical, scalable MUAD solution and demonstrates strong generalization to diverse, real-world data while providing clear ablation-driven insights into component effectiveness.

Abstract

Multi-class unsupervised anomaly detection (MUAD) has garnered growing research interest, as it seeks to develop a unified model for anomaly detection across multiple classes, i.e., eliminating the need to train separate models for distinct objects and thereby saving substantial computational resources. Under the MUAD setting, while advanced Transformer-based architectures have brought significant performance improvements, identity shortcuts persist: they directly copy inputs to outputs, narrowing the gap in reconstruction errors between normal and abnormal cases, and thereby making the two harder to distinguish. Therefore, we propose ShortcutBreaker, a novel unified feature-reconstruction framework for MUAD tasks, featuring two key innovations to address the issue of shortcuts. First, drawing on matrix rank inequality, we design a low-rank noisy bottleneck (LRNB) to project highdimensional features into a low-rank latent space, and theoretically demonstrate its capacity to prevent trivial identity reproduction. Second, leveraging ViTs global modeling capability instead of merely focusing on local features, we incorporate a global perturbation attention to prevent information shortcuts in the decoders. Extensive experiments are performed on four widely used anomaly detection benchmarks, including three industrial datasets (MVTec-AD, ViSA, and Real-IAD) and one medical dataset (Universal Medical). The proposed method achieves a remarkable image-level AUROC of 99.8%, 98.9%, 90.6%, and 87.8% on these four datasets, respectively, consistently outperforming previous MUAD methods across different scenarios.

Figures (6)

• Figure 1: Performance comparison in terms of image-level AUC on MVTec, ViSA , Universal Medical and Real-IAD .
• Figure 2: The visual patterns in the medical field are richer than those in the industrial field. (a) Visualization results, (b) T-SNE plots van2008visualizing, and (c) distributions of feature distances for the Universal Medical and ViSA datasets zou2022spot are presented. These experimental results help observe the diversity of each dataset. For T-SNE plots, we extract the final features from the pre-trained ResNet-50 he2016deep. Moreover, we use LPIPs zhang2018unreasonable to compare feature distances between individual image pairs.
• Figure 3: Flowchart of our ShortcutBreaker: (a) the overall structure of our proposed method; (b) the structure of the Low-Rank Noisy Bottleneck (LRNB); (c) the structure of the Global Perturbation Attention (GPA), which consists of GRD (global redistribution) and GSM (global-self masking); and (d) the training and inference pipelines.
• Figure 4: Comparison of sigmoid and softmax Functions on a 2D attention map with Gaussian distribution, (a), (b), (c) are the 3D plot of original data, after softmax and sigmoid respectively, (d), (e), (f) are the corresponding heatmaps.
• Figure 5: The flowchart of Global-Self Masking operation. AM: attention map and ($\cdot$) indicate element-wise multiplication.