Omni-frequency Channel-selection Representations for Unsupervised Anomaly Detection (2203.00259v2)

Abstract: Density-based and classification-based methods have ruled unsupervised anomaly detection in recent years, while reconstruction-based methods are rarely mentioned for the poor reconstruction ability and low performance. However, the latter requires no costly extra training samples for the unsupervised training that is more practical, so this paper focuses on improving this kind of method and proposes a novel Omni-frequency Channel-selection Reconstruction (OCR-GAN) network to handle anomaly detection task in a perspective of frequency. Concretely, we propose a Frequency Decoupling (FD) module to decouple the input image into different frequency components and model the reconstruction process as a combination of parallel omni-frequency image restorations, as we observe a significant difference in the frequency distribution of normal and abnormal images. Given the correlation among multiple frequencies, we further propose a Channel Selection (CS) module that performs frequency interaction among different encoders by adaptively selecting different channels. Abundant experiments demonstrate the effectiveness and superiority of our approach over different kinds of methods, e.g., achieving a new state-of-the-art 98.3 detection AUC on the MVTec AD dataset without extra training data that markedly surpasses the reconstruction-based baseline by +38.1 and the current SOTA method by +0.3. Source code is available at https://github.com/zhangzjn/OCR-GAN.

• The paper introduces the OCR-GAN network that decouples omni-frequency channels to boost reconstruction-based anomaly detection.
• The method employs frequency decoupling and adaptive channel selection, enabling superior performance with no extra training data.
• Results show that OCR-GAN achieves a 98.3 AUC on MVTec AD, marking a significant advance over existing techniques.

Omni-frequency Channel-selection Representations for Unsupervised Anomaly Detection

The paper "Omni-frequency Channel-selection Representations for Unsupervised Anomaly Detection" presents a novel approach named Omni-frequency Channel-selection Reconstruction (OCR-GAN) network. This method aims to improve the unsupervised anomaly detection task, specifically focusing on the reconstruction-based technique. This approach is crucial due to its ability to operate without the need for costly extra training samples, positioning it as a practical solution in various real-world applications.

Reconstruction-based Anomaly Detection

Most existing unsupervised anomaly detection techniques work on either density-based or classification-based methods, which usually deliver high performance but require pre-trained models and additional training data, thus becoming less practical. Traditionally, reconstruction-based methods have underachieved due to their poor reconstruction capability. However, these methods offer an advantage by needing no extra data. The paper targets the reconstruction-based approach and proposes the OCR-GAN network to significantly enhance performance within this framework.

Omni-frequency Channel-selection Mechanism

The core innovation in OCR-GAN comes from the introduction of frequency domain analysis and interaction among different frequency channels. The method leverages the difference in frequency distribution between normal and abnormal sensory images. This difference is crucial since normal and abnormal images have distinctive frequency signatures, which the OCR-GAN exploits for improved detection.

Frequency Decoupling (FD) Module

The FD module decomposes an image into multiple frequency components, allowing parallel reconstruction of omni-frequency images. This decomposition is based on the observation that different frequencies encapsulate varied types of information (e.g., texture at high frequencies, semantics at low frequencies). By enabling separate frequency-wise reconstruction, the model can better focus on the details that differentiate normal from abnormal conditions.

Channel Selection (CS) Module

To further enhance the interaction among these distinct frequency branches, a CS module is introduced, promoting adaptive channel selection based on different frequency interactions. This module effectively maps correlations among channels in various frequency branches, augmenting the reconstruction process by incorporating omni-frequency features adaptively.

Experimental Results and Implications

OCR-GAN experimentally demonstrates superior performance across a variety of benchmark datasets, including MVTec AD, DAGM, and KolektorSDD, as well as the CIFAR-10 for semantic AD task. Specifically, OCR-GAN surpassed current state-of-the-art methods on the MVTec AD dataset with a notable AUC of 98.3 without requiring extra training data, marking a significant improvement over previous classical reconstruction-based methods. The ability to achieve this high AUC without additional data emphasizes the model's practicality for real-world applications where training data is limited or expensive to obtain.

Future Directions

The paper suggests that reconstruction-based methods can achieve competitive scores through careful design and frequency domain exploitation, opening avenues for further exploration of lightweight models and frequency-sensitive anomaly detection strategies. Future work might address extending the omni-frequency reconstruction concept across other domains and further refining the lightweight model design to balance efficiency and efficacy comprehensively.

In summary, the OCR-GAN offers a crucial step forward in the reconstruction-based anomaly detection arena, providing a practical and high-performing solution that includes decoupling and interacting frequency components efficiently for enhanced anomaly detection.