RealNet is a novel self-supervised feature reconstruction network designed for anomaly detection in industrial images. It addresses the challenges of synthesizing realistic and diverse anomaly samples, as well as reducing feature redundancy and pre-training bias. The key contributions of RealNet include: 1. **Strength-controllable Diffusion Anomaly Synthesis (SDAS)**: A diffusion process-based method that generates samples with varying anomaly strengths, mimicking real-world anomaly distributions. 2. **Anomaly-aware Features Selection (AFS)**: A method that selects representative and discriminative pre-trained feature subsets to improve anomaly detection performance while controlling computational costs. 3. **Reconstruction Residuals Selection (RRS)**: A strategy that adaptively selects discriminative residuals to enhance the identification of anomalous regions across multiple granularities. RealNet is evaluated on four benchmark datasets (MVTec-AD, MPDD, BTAD, and VisA), demonstrating significant improvements in both Image AUROC and Pixel AUROC compared to state-of-the-art methods. The code, data, and models are available at <https://github.com/cnulab/RealNet>. - **SDAS**: Generates diverse, near-natural distribution anomalous images. - **AFS**: Refines pre-trained features for dimensionality reduction and bias elimination. - **RRS**: Selects discriminative residuals for comprehensive anomaly identification. - **MVTec-AD**: Achieves 99.65% Image AUROC, 99.03% Pixel AUROC, and 93.07% PRO score. - **MPDD**: Improves by 2.88% in Image AUROC over DTD. - **BTAD**: Achieves 96.1% Image AUROC and 97.9% Pixel AUROC. - **VisA**: Achieves 97.8% Image AUROC and 98.8% Pixel AUROC. - **AFS and RRS**: Significantly improve performance. - **Anomaly Strength ($s$)**: Uniformly sampling $s$ within a specific range enhances performance. - **RRS Modes and Retention Ratios**: Max&Avg mode outperforms Max and Avg modes in detecting anomalies across various scales. RealNet effectively leverages large-scale pre-trained models for anomaly detection, providing a flexible foundation for future research in this area.RealNet is a novel self-supervised feature reconstruction network designed for anomaly detection in industrial images. It addresses the challenges of synthesizing realistic and diverse anomaly samples, as well as reducing feature redundancy and pre-training bias. The key contributions of RealNet include: 1. **Strength-controllable Diffusion Anomaly Synthesis (SDAS)**: A diffusion process-based method that generates samples with varying anomaly strengths, mimicking real-world anomaly distributions. 2. **Anomaly-aware Features Selection (AFS)**: A method that selects representative and discriminative pre-trained feature subsets to improve anomaly detection performance while controlling computational costs. 3. **Reconstruction Residuals Selection (RRS)**: A strategy that adaptively selects discriminative residuals to enhance the identification of anomalous regions across multiple granularities. RealNet is evaluated on four benchmark datasets (MVTec-AD, MPDD, BTAD, and VisA), demonstrating significant improvements in both Image AUROC and Pixel AUROC compared to state-of-the-art methods. The code, data, and models are available at <https://github.com/cnulab/RealNet>. - **SDAS**: Generates diverse, near-natural distribution anomalous images. - **AFS**: Refines pre-trained features for dimensionality reduction and bias elimination. - **RRS**: Selects discriminative residuals for comprehensive anomaly identification. - **MVTec-AD**: Achieves 99.65% Image AUROC, 99.03% Pixel AUROC, and 93.07% PRO score. - **MPDD**: Improves by 2.88% in Image AUROC over DTD. - **BTAD**: Achieves 96.1% Image AUROC and 97.9% Pixel AUROC. - **VisA**: Achieves 97.8% Image AUROC and 98.8% Pixel AUROC. - **AFS and RRS**: Significantly improve performance. - **Anomaly Strength ($s$)**: Uniformly sampling $s$ within a specific range enhances performance. - **RRS Modes and Retention Ratios**: Max&Avg mode outperforms Max and Avg modes in detecting anomalies across various scales. RealNet effectively leverages large-scale pre-trained models for anomaly detection, providing a flexible foundation for future research in this area.