[Paper Review] RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection

Ximiao Zhang, Min Xu+1 more|arXiv (Cornell University)|Mar 9, 2024

Anomaly Detection Techniques and Applications4 citations

TL;DR

RealNet proposes a self-supervised anomaly detection framework that combines strength-controllable diffusion-based anomaly synthesis (SDAS), adaptive feature selection (AFS), and reconstruction residual selection (RRS) to improve detection performance while reducing computational cost. It achieves state-of-the-art results, including 96.3% Image AUROC on MVTec-AD and 97.8% on VisA, outperforming prior methods without dataset-specific tuning.

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.

Motivation & Objective

• To address the limitations of existing self-supervised anomaly detection methods in generating realistic and diverse synthetic anomalies.
• To reduce feature redundancy and pre-training bias in large-scale pre-trained models used for anomaly detection.
• To improve detection performance and recall by adaptively selecting discriminative reconstruction residuals across multiple scales.
• To develop a unified, flexible framework that optimizes computational cost while maintaining high detection accuracy.
• To provide a new benchmark synthetic dataset (SIA) for future self-supervised anomaly detection research.

Proposed method

• Introduces Strength-controllable Diffusion Anomaly Synthesis (SDAS), a diffusion-based method that generates synthetic anomalies with controllable strength by perturbing normal samples in low-density regions.
• Proposes Anomaly-aware Features Selection (AFS), a method that adaptively selects optimal pre-trained feature subsets for each anomaly category to enhance discriminability and reduce redundancy.
• Develops Reconstruction Residuals Selection (RRS), which adaptively filters out non-discriminative residuals across multi-level features to improve localization accuracy.
• Employs a multi-scale reconstruction framework that leverages features from multiple layers of a pre-trained CNN, with residual selection based on anomaly relevance.
• Uses a discriminator to score anomalies based on selected reconstruction residuals, enabling end-to-end training with minimal hyperparameter tuning.
• Introduces a retention ratio $P = r / \sum_{k=1}^{K} m_k$ to control the proportion of residuals retained, with Max & Avg mode showing superior robustness.

Experimental results

Research questions

• RQ1 Can diffusion-based synthesis generate realistic and diverse synthetic anomalies that better match real-world anomaly distributions?
• RQ2 How can pre-trained features be adaptively selected to improve detection performance while minimizing computational cost?
• RQ3 Does adaptive selection of reconstruction residuals improve detection recall and localization accuracy across different anomaly scales?
• RQ4 Can a unified framework achieve state-of-the-art performance across multiple benchmarks without dataset-specific tuning?
• RQ5 How does the strength of synthetic anomalies affect the generalization and robustness of the final model?

Key findings

• RealNet achieves 96.3% Image AUROC on MVTec-AD, surpassing the previous SOTA (CFlow) by 10.2% without any dataset-specific tuning.
• On the VisA dataset, RealNet achieves 97.8% Image AUROC and 98.8% Pixel AUROC, demonstrating strong generalization on complex, multi-object anomalies.
• On BTAD, RealNet achieves 96.1% Image AUROC and 97.9% Pixel AUROC, showing consistent performance across diverse industrial datasets.
• The Max & Avg mode of RRS achieves the best robustness and performance, outperforming Max or Avg alone by effectively filtering non-discriminative residuals.
• Ablation studies confirm that AFS significantly outperforms random feature selection methods like RDR and RLPR, highlighting its effectiveness in feature subset selection.
• Optimal anomaly strength $s$ in SDAS lies within a specific range—too low causes false positives, too high causes distribution shift, with uniform sampling across $s$ yielding the most robust results.