Progressive Boundary Guided Anomaly Synthesis for Industrial Anomaly Detection

TL;DR

This work tackles unsupervised industrial anomaly detection by proposing PBAS, a boundary-guided anomaly synthesis framework that does not rely on auxiliary textures. PBAS comprises Approximate Boundary Learning (ABL) to compact normal features on a hypersphere, Anomaly Feature Synthesis (AFS) to generate directional anomalies along rays from the center, and Refined Boundary Optimization (RBO) to sharpen the boundary via a discriminative network. The method achieves state-of-the-art results on MVTec AD, VisA, and MPDD with rapid inference, demonstrating strong generalization to diverse industrial defects and high localization accuracy. By removing predefined anomaly properties and enabling self-adaptive, directional synthesis, PBAS offers a practical and efficient approach for real-world quality inspection.

Abstract

Unsupervised anomaly detection methods can identify surface defects in industrial images by leveraging only normal samples for training. Due to the risk of overfitting when learning from a single class, anomaly synthesis strategies are introduced to enhance detection capability by generating artificial anomalies. However, existing strategies heavily rely on anomalous textures from auxiliary datasets. Moreover, their limitations in the coverage and directionality of anomaly synthesis may result in a failure to capture useful information and lead to significant redundancy. To address these issues, we propose a novel Progressive Boundary-guided Anomaly Synthesis (PBAS) strategy, which can directionally synthesize crucial feature-level anomalies without auxiliary textures. It consists of three core components: Approximate Boundary Learning (ABL), Anomaly Feature Synthesis (AFS), and Refined Boundary Optimization (RBO). To make the distribution of normal samples more compact, ABL first learns an approximate decision boundary by center constraint, which improves the center initialization through feature alignment. AFS then directionally synthesizes anomalies with more flexible scales guided by the hypersphere distribution of normal features. Since the boundary is so loose that it may contain real anomalies, RBO refines the decision boundary through the binary classification of artificial anomalies and normal features. Experimental results show that our method achieves state-of-the-art performance and the fastest detection speed on three widely used industrial datasets, including MVTec AD, VisA, and MPDD. The code will be available at: https://github.com/cqylunlun/PBAS.

Figures (13)

• Figure 1: Overview of the existing and proposed anomaly synthesis strategies. (a) Image-level anomaly synthesis strategies heavily rely on the predefined textures from auxiliary datasets. (b) Feature-level anomaly synthesis strategies utilize Gaussian noise with a fixed scale in random directions. (c) Our proposed method directionally synthesizes feature-level anomalies with flexible scales and without predefined properties.
• Figure 2: Conceptual illustration of our proposed PBAS. (a) The initial distribution of normal samples in the feature space is relatively dispersed. (b) Through the learning of ABL, normal features are projected into a compact hypersphere. (c) Through the synthesis of AFS, artificial anomalies are generated from normal features by the hypersphere distribution. (d) Through the optimization of RBO, the decision boundary is further refined by the discriminative network.
• Figure 3: Schematic of the proposed PBAS. (a) Approximate Boundary Learning (ABL) learns an approximate boundary of normal features through the improved center constraint. (b) Anomaly Feature Synthesis (AFS) synthesizes anomaly features based on the hypersphere distribution of the ABL output. (c) Refined Boundary Optimization (RBO) further refines the boundary through discriminative learning of the AFS output. The training stage is depicted with solid and dashed arrows, while the inference stage is indicated by solid arrows.
• Figure 4: Comparison of center initialization methods. (a) The traditional method obtains center feature by the average feature of entire training set through a single feature extractor. (b) Our method obtains center feature by feature alignment with iterative updates through a pair of feature extractor and projector.
• Figure 5: Mechanism of the self-adaptive generation by AFS and progressive boundary guidance by ABL and RBO.