Few-Shot Anomaly Detection via Category-Agnostic Registration Learning

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional anomaly detection (AD) methods necessitate category-specific models, leading to high computational costs and inefficiency.
• Existing paradigms struggle with real-world applications due to the need for extensive, labeled datasets for each new category.
• Human-like anomaly detection, comparing novel instances to known normal examples, offers an alternative paradigm.

Purpose of the Study:

• To propose a novel few-shot anomaly detection (FSAD) framework that overcomes the limitations of category-specific models.
• To develop a computationally efficient and generalizable AD solution applicable to novel categories without fine-tuning.
• To enable self-supervised learning of category-agnostic representations for robust anomaly identification.

Main Methods:

• Leveraged image registration as a proxy task for self-supervised, category-agnostic representation learning using normal images.
• Developed a framework where a test image is compared to a small support set of normal images from the same category.
• Utilized aligned features from registered images to identify anomalies, enabling generalization to unseen categories.

Main Results:

• Achieved state-of-the-art (SOTA) performance on FSAD benchmarks, improving results by 11.3% on MVTec and 8.3% on MPDD.
• Demonstrated the model's ability to generalize to novel test categories without requiring any model fine-tuning.
• Validated the effectiveness of the proposed registration-based, category-agnostic approach for efficient anomaly detection.

Conclusions:

• The proposed FSAD framework offers a significant advancement in anomaly detection by eliminating the need for category-specific models.
• This method provides a computationally efficient and highly generalizable solution for real-world anomaly detection tasks.
• The framework represents the first FSAD method capable of handling novel categories without fine-tuning, paving the way for broader applications.