Improving anomaly detection with foundation-model synthesis and wavelet-domain attention.

Area of Science:

• Industrial anomaly detection
• Machine Learning
• Computer Vision

Background:

• Industrial anomaly detection is challenged by limited anomalous data and complex anomaly types.
• Existing methods often require extensive training data or struggle with diverse anomalies.

Purpose of the Study:

• To develop a foundation model-based anomaly synthesis pipeline (FMAS) for generating realistic anomalous samples.
• To introduce a Wavelet Domain Attention Module (WDAM) for enhanced anomaly feature extraction in the frequency domain.
• To improve the sensitivity and efficiency of industrial anomaly detection systems.

Main Methods:

• FMAS generates synthetic anomalies without fine-tuning or class-specific training.
• WDAM utilizes adaptive sub-band processing in the wavelet domain to focus on anomaly-related features.
• The proposed methods were evaluated on MVTec AD and VisA datasets.

Main Results:

• The FMAS pipeline successfully generates highly realistic anomalous samples.
• WDAM significantly enhances anomaly feature extraction by exploiting frequency-domain characteristics.
• The combined approach demonstrates substantial performance gains over existing baselines.
• WDAM functions as an effective plug-and-play module.

Conclusions:

• The proposed FMAS and WDAM offer a powerful solution for industrial anomaly detection, particularly when anomalous data is scarce.
• WDAM improves detection sensitivity and maintains computational efficiency.
• These advancements contribute to more robust and effective automated inspection systems.