RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (2403.05897v1)

Abstract: Self-supervised feature reconstruction methods have shown promising advances in industrial image anomaly detection and localization. Despite this progress, these methods still face challenges in synthesizing realistic and diverse anomaly samples, as well as addressing the feature redundancy and pre-training bias of pre-trained feature. In this work, we introduce RealNet, a feature reconstruction network with realistic synthetic anomaly and adaptive feature selection. It is incorporated with three key innovations: First, we propose Strength-controllable Diffusion Anomaly Synthesis (SDAS), a diffusion process-based synthesis strategy capable of generating samples with varying anomaly strengths that mimic the distribution of real anomalous samples. Second, we develop Anomaly-aware Features Selection (AFS), a method for selecting representative and discriminative pre-trained feature subsets to improve anomaly detection performance while controlling computational costs. Third, we introduce Reconstruction Residuals Selection (RRS), a strategy that adaptively selects discriminative residuals for comprehensive identification of anomalous regions across multiple levels of granularity. We assess RealNet on four benchmark datasets, and our results demonstrate significant improvements in both Image AUROC and Pixel AUROC compared to the current state-o-the-art methods. The code, data, and models are available at https://github.com/cnulab/RealNet.

• The paper introduces RealNet, a framework that integrates diffusion-based anomaly synthesis (SDAS), anomaly-aware feature selection (AFS), and residual selection (RRS) for enhanced detection.
• RealNet generates realistic synthetic anomalies while reducing feature redundancy to improve both image and pixel-level detection performance.
• Evaluations on benchmark datasets demonstrate superior anomaly localization accuracy, paving the way for future generative approaches in industrial imaging.

Overview of RealNet: A Feature Selection Network for Anomaly Detection

The paper introduces RealNet, a novel framework for industrial image anomaly detection, focusing on the challenges of realistic anomaly synthesis and efficient feature selection. RealNet advances the field by incorporating Strength-controllable Diffusion Anomaly Synthesis (SDAS), Anomaly-aware Features Selection (AFS), and Reconstruction Residuals Selection (RRS). The paper presents a comprehensive evaluation on benchmark datasets, showcasing RealNet's performance improvements over existing methods.

RealNet's architecture is grounded on three pivotal innovations:

1. Strength-controllable Diffusion Anomaly Synthesis (SDAS): SDAS leverages a diffusion-based process to generate synthetic anomalies with controllable strengths. This approach addresses the limitations of existing methods that rely on data augmentation or external datasets, offering improved realism and diversity in synthetic anomalies. The implementation of SDAS is grounded in modeling the anomalies as perturbations in low probability density regions of the normal data distribution.
2. Anomaly-aware Features Selection (AFS): AFS systematically identifies the most relevant subset of pre-trained feature maps, enhancing the discriminatory capacity of RealNet while managing computational costs. This process significantly mitigates the dimensionality and redundancy issues commonly associated with high-dimensional, pre-trained CNN features. The feature selection is conducted in a self-supervised manner, utilizing synthetic anomalies generated by SDAS.
3. Reconstruction Residuals Selection (RRS): RRS selectively processes reconstruction residuals to enhance localization accuracy across varying levels of granularity. By adaptively discarding less informative residuals, RRS improves RealNet's capability to identify and recall anomalous regions. The integration of RRS into the anomaly detection pipeline contributes to a tangible uplift in both Image AUROC and Pixel AUROC metrics, as evidenced by the experimental results.

The empirical evaluations conducted on datasets such as MVTec-AD, MPDD, BTAD, and VisA, demonstrate the superiority of RealNet in terms of image and pixel-level anomaly detection and localization. The inclusion of the Synthetic Industrial Anomaly Dataset (SIA), generated by SDAS, further facilitates the advancement of self-supervised anomaly detection approaches.

Implications and Future Directions

The research embodies a significant step towards improving the practical applicability of anomaly detection systems in industrial contexts. By addressing both the synthesis of realistic anomalies and the efficient use of pre-trained features, RealNet enhances the generalization capability of detection models to real-world data variations.

From a theoretical perspective, the successful integration of diffusion models in anomaly synthesis highlights the potential for leveraging generative frameworks in anomaly detection tasks. This approach opens up future research avenues exploring different generative architectures and their applications in creating highly plausible anomaly scenarios.

Additionally, the adaptive selection mechanisms developed for feature maps (AFS) and reconstruction residuals (RRS) in RealNet emphasize the importance of targeted, efficient model designs for handling large-scale pre-trained networks, which could inspire further research into optimization techniques within other domains of computer vision.

Conclusively, the RealNet framework is a robust contribution to the field of anomaly detection, with practical implications for improving safety and quality assurance across various industrial sectors. Future work could explore expanding RealNet's applicability to other anomaly detection scenarios outside of industrial imaging, such as cybersecurity and medical imaging.