"Stones From Other Hills Can Polish Jade": Zero-Shot Anomaly Synthesis via Cross-Domain Anomaly Injection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Industrial image anomaly detection (IAD) is crucial but hindered by the rarity of domain-specific anomalies.
• Existing zero-shot anomaly synthesis (ZSAS) methods struggle with authenticity or require extensive training.
• The scarcity of in-domain anomalies limits the effectiveness of current IAD techniques.

Purpose of the Study:

• To propose a novel, training-free paradigm for authentic zero-shot anomaly synthesis (ZSAS) in industrial image anomaly detection (IAD).
• To address the challenge of rare domain-specific anomalies by utilizing abundant cross-domain anomalies.
• To develop a pragmatic and effective solution for enhancing IAD systems.

Main Methods:

• Cross-domain Anomaly Injection (CAI): A novel method exploiting cross-domain anomalies for training-free ZSAS.
• Domain-agnostic Anomaly Dataset (DAAD): The first dataset of its kind, providing abundant real anomaly patterns for ZSAS.
• CAI-guided Diffusion Mechanism: Enhances anomaly synthesis by breaking quantity limits of real anomalies.

Main Results:

• The proposed CAI method achieves highly authentic ZSAS without any training.
• The DAAD dataset offers a rich source of diverse, real-world anomaly patterns.
• The CAI-guided Diffusion Mechanism enables unlimited anomaly synthesis, significantly boosting ZSAS capabilities.
• Head-to-head comparisons demonstrate superior performance over existing ZSAS solutions.

Conclusions:

• The novel ZSAS paradigm effectively generates authentic pseudo anomalies for IAD using cross-domain data.
• The developed methods (CAI, DAAD, Diffusion Mechanism) offer a pragmatic and scalable solution for IAD.
• This approach overcomes the limitations of rare domain-specific anomalies, advancing the field of industrial image analysis.