A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization (2407.09359v1)

Abstract: Anomaly synthesis strategies can effectively enhance unsupervised anomaly detection. However, existing strategies have limitations in the coverage and controllability of anomaly synthesis, particularly for weak defects that are very similar to normal regions. In this paper, we propose Global and Local Anomaly co-Synthesis Strategy (GLASS), a novel unified framework designed to synthesize a broader coverage of anomalies under the manifold and hypersphere distribution constraints of Global Anomaly Synthesis (GAS) at the feature level and Local Anomaly Synthesis (LAS) at the image level. Our method synthesizes near-in-distribution anomalies in a controllable way using Gaussian noise guided by gradient ascent and truncated projection. GLASS achieves state-of-the-art results on the MVTec AD (detection AUROC of 99.9\%), VisA, and MPDD datasets and excels in weak defect detection. The effectiveness and efficiency have been further validated in industrial applications for woven fabric defect detection. The code and dataset are available at: \url{https://github.com/cqylunlun/GLASS}.

• The paper introduces GLASS, a unified framework combining global feature-level and local image-level synthesis via gradient ascent to enhance industrial anomaly detection.
• It leverages Global Anomaly Synthesis (GAS) and Local Anomaly Synthesis (LAS) to generate diverse anomaly textures and tighten classification boundaries.
• Experimental results, including a 99.9% AUROC on MVTec AD, demonstrate its superiority over state-of-the-art methods.

An Exploration of a Unified Anomaly Synthesis Strategy for Industrial Detection

In the field of industrial anomaly detection and localization, challenges persist due to the scarcity of annotated data and the high cost of obtaining pixel-level annotations for defects. The paper "A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization" presents a novel approach to overcoming these challenges through an overview-based technique.

Methodology

The authors propose a unified framework named Global and Local Anomaly co-Synthesis Strategy (GLASS), which synthesizes anomalies at both the feature and image levels. This dual-layer synthesis employs the Global Anomaly Synthesis (GAS) and Local Anomaly Synthesis (LAS) methodologies. The method is designed to maximize anomaly coverage and control using gradient ascent alongside truncated projection techniques.

1. Global Anomaly Synthesis (GAS): GAS operates at the feature level, controlling anomaly synthesis by using Gaussian noise augmented by gradient ascent. This approach allows for a tighter classification boundary, enhancing the detection of weak defects adjacent to normal samples.
2. Local Anomaly Synthesis (LAS): At the image level, LAS creates a variety of anomaly textures by overlaying textures onto normal samples, enhancing the range and variety of synthesized anomalies.

Key Results

The paper reports superior performance of the proposed GLASS framework compared to state-of-the-art (SOTA) methods across several datasets including MVTec AD, VisA, MPDD, and a newly introduced WFDD dataset. Noteworthy results include a 99.9% AUROC for anomaly detection on the MVTec AD dataset, demonstrating its efficacy in both strong and weak defect detection scenarios.

Implications and Future Directions

The significant efficacy of the GLASS framework in both industrial and complex environments suggests that synthesis-based anomaly detection could be a viable solution in applications with limited defect data and high real-time processing demands. The approach leverages both feature-level and image-level synthesis, providing a comprehensive solution that addresses variability in defect presentation.

The synthesis strategy presents interesting implications for future AI developments. The usage of gradient ascent for anomaly synthesis suggests a potential for adaptability of the synthesis parameters, aligning the generation of anomalies with specific industrial defect distributions. Future research could explore extending this synthesis framework to logical anomaly detection, expanding its application domain beyond structural anomalies. Moreover, eliminating the dependency on external texture datasets for image-level anomaly synthesis might further streamline the framework, enhancing its integration within industrial settings.

In conclusion, the paper marks a significant contribution to the world of unsupervised anomaly detection by providing a comprehensive framework capable of detecting a broad spectrum of defects efficiently. While it does not fundamentally challenge the existing paradigms, it offers a profound enhancement in the detection and localization of anomalies, pushing forward the capabilities of industrial inspection systems.