RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection

Ximiao Zhang,Min Xu,Xiuzhuang Zhou

2024-03-09

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 this https URL.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to address several key issues in industrial image anomaly detection:

1. **Synthesizing realistic and diverse anomaly samples**: Existing self-supervised feature reconstruction methods face challenges in generating realistic and diverse anomaly samples, especially when it comes to generating complex structural anomalies and unknown anomaly categories.

2. **Feature redundancy and pre-training bias**: Existing methods encounter the problem of feature redundancy when dealing with high-dimensional features, and pre-trained networks introduce biases from classification tasks, which limit the performance of large-scale pre-trained networks in anomaly detection.

To address these issues, the paper proposes the RealNet framework, which includes three main innovations:
- **Strength-controllable Diffusion Anomaly Synthesis (SDAS)**: A synthesis strategy based on the diffusion process that can generate anomaly samples with different intensities, simulating the distribution of real anomaly samples.
- **Anomaly-aware Feature Selection (AFS)**: A method for selecting representative subsets from pre-trained networks to improve anomaly detection performance while controlling computational costs.
- **Reconstruction Residual Selection (RRS)**: A strategy for selecting discriminative reconstruction residuals at different granularity levels to comprehensively identify anomaly regions.

With these three modules, RealNet achieves significantly better performance than current state-of-the-art methods on four benchmark datasets (MVTec-AD, MPDD, BTAD, and VisA). Additionally, the paper provides a generated Synthetic Industrial Anomaly dataset (SIA), which includes 360,000 anomaly images from 36 industrial product categories, for further research and application.