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

TL;DR

RealNet tackles industrial anomaly detection by jointly addressing realistic anomaly synthesis and efficient, discriminative feature reconstruction. It combines Strength-controllable Diffusion Anomaly Synthesis (SDAS) with Anomaly-aware Features Selection (AFS) and Reconstruction Residuals Selection (RRS) to leverage large-scale pre-trained features while reducing redundancy and bias. The Synthetic Industrial Anomaly Dataset (SIA) enables self-supervised learning of diverse anomalies, and RealNet achieves state-of-the-art Image AUROC, Pixel AUROC, and PRO across four benchmarks, with practical inference speed. This framework offers a flexible, scalable approach for robust anomaly detection in real-world industrial settings.

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.

Figures (16)

• Figure 1: SDAS generates anomaly images using only normal images. The example images are sourced from the MVTec-AD dataset bergmann2019mvtec.
• Figure 2: The pipeline of our RealNet consists of three core components: Strength-controllable Diffusion Anomaly Synthesis (SDAS), Anomaly-aware Features Selection (AFS), and Reconstruction Residuals Selection (RRS). 1) SDAS enables the synthesis of diverse, near-natural distribution anomalous images. 2) AFS refines features extracted by large-scale pre-trained CNN for dimensionality reduction. Refined features are reconstructed into corresponding normal image features by a set of reconstruction networks. 3) RRS selects reconstruction residuals most likely to identify anomalies, which are then fed into a discriminator for anomaly detection and localization.
• Figure 3: Anomaly image examples generated with different synthesis methods. (a) Examples generated using SDAS with different anomaly strengths $s$. (b) Examples featuring local anomaly regions generated by various anomaly synthesis methods.
• Figure 4: Comparison of SIA with DTD cimpoi2014describing and CutPaste li2021cutpaste on the MPDD dataset jezek2021deep, employing Image AUROC (%), Pixel AUROC (%), and PRO (%) as evaluation metrics.
• Figure 5: Qualitative results of RealNet on the MVTec-AD dataset bergmann2019mvtec. Within each group, from left to right, are the anomaly image, ground-truth, and predicted anomaly score.