Research map / Label-efficient and robust learning / OOD and anomaly detection

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

CVPR 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 198 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RealNet, a feature reconstruction anomaly detector combining diffusion-based strength-controllable anomaly synthesis (SDAS), anomaly-aware feature selection (AFS), and reconstruction residual selection (RRS), plus a released 360k-image synthetic dataset (SIA).

Problem

Industrial image anomaly detection and localization with self-supervised feature reconstruction suffers from unrealistic synthetic anomalies, feature redundancy, and pre-training bias of pre-trained features.

Approach

Generate anomaly images by perturbing DDPM reverse sampling with a strength parameter s so samples lie in low-density regions near normal data; select a subset of pre-trained feature maps per layer using synthetic anomalies (AFS); and keep only the top-K reconstruction residuals via global max/avg pooling (RRS) before a discriminator.

Key results

• On MVTec-AD, RealNet trained with SIA reaches 99.65% Image AUROC, 99.03% Pixel AUROC and 93.07% PRO, outperforming listed baselines such as PatchCore (99.1) and UniAD (96.6) in Image AUROC.
• On MPDD, RealNet achieves 96.3% Image AUROC versus 86.1% for CFlow, the best listed baseline.
• Ablation on MVTec-AD: removing RRS drops Image AUROC from 99.65 to 99.39 with RRS-only... combining AFS and RRS gives the best results; RealNet runs at 31.93 FPS on an RTX 3090 with 4GB GPU memory.

Measured effects reported in the paper

Surprising findings

• Setting anomaly strength s=0 (generating normal-like images) performs worse than sampling s in [0.1,0.2], as blending may introduce false-positive anomalies.

Limitations

• Reliance on synthetic anomalies generated from normal images may not cover all real anomaly types.
• Feature selection and reconstruction depend on the chosen backbone and per-layer dimension hyperparameters.

Experimental setup

Datasets: MVTec-AD MPDD BTAD VisA DTD SIA

Benchmarks: MVTec-AD MPDD BTAD VisA

Metrics: Image AUROC Pixel AUROC PRO FPS

Baselines: PatchCore SimpleNet FastFlow DRAEM+SSPCAB DSR UniAD RD++ DeSTSeg DiffAD CFlow PaDiM SPADE

Builds on: DDPM DRAEM CutPaste DTD texture dataset Perlin noise anomaly generation

Closest papers by meaning

• A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.88)
• Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
• AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.87)
• A Comprehensive Augmentation Framework for Anomaly Detection (AAAI 2024 · similarity 0.87)
• ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.87)
• FAST: Foreground‑aware Diffusion with Accelerated Sampling Trajectory for Segmentation‑oriented Anomaly Synthesis (NeurIPS 2025 · similarity 0.86)