Omni-Frequency Channel-Selection Representations for Unsupervised Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Unsupervised anomaly detection commonly uses density-based and classification-based methods.
• Reconstruction-based methods are often overlooked due to performance limitations.
• However, reconstruction-based methods offer practical advantages by avoiding costly extra training samples.

Purpose of the Study:

• To enhance reconstruction-based methods for unsupervised anomaly detection.
• To introduce a novel Omni-frequency Channel-selection Reconstruction (OCR-GAN) network.
• To address sensory anomaly detection using a frequency-based perspective.

Main Methods:

• Proposed a Frequency Decoupling (FD) module to separate input images into distinct frequency components.
• Modeled reconstruction as parallel restorations across multiple frequencies, exploiting differences between normal and abnormal image frequency distributions.
• Introduced a Channel Selection (CS) module for adaptive frequency interaction among encoders.

Main Results:

• Achieved a state-of-the-art 98.3 detection AUC on the MVTec AD dataset.
• Significantly outperformed the reconstruction-based baseline by +38.1 AUC.
• Surpassed the current state-of-the-art method by +0.3 AUC.
• Demonstrated effectiveness and superiority over various anomaly detection methods.

Conclusions:

• The OCR-GAN network effectively improves reconstruction-based unsupervised anomaly detection.
• Analyzing image frequencies and their interactions is crucial for enhanced detection.
• The proposed method offers a practical and high-performing solution for anomaly detection tasks.