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

Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection

CVPR 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 50 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 INP-Former, which extracts intrinsic normal prototypes from a single test image via a learnable cross-attention INP Extractor and reconstructs normal tokens with an INP-guided decoder for universal (single-, multi-class, few-shot) anomaly detection.

Problem

Anomaly detection relies on comparing test images to normal references from a training set, but appearance and position variations cause misaligned normality, limiting accuracy in single-class, multi-class, and few-shot settings.

Approach

Instead of pre-stored normal prototypes from the training set, learn M intrinsic normal prototypes as linear combinations of tokens within the test image itself, use them as keys/values in an INP-guided decoder that reconstructs only normal tokens, and score anomalies by reconstruction error.

Key results

• Multi-class AD: image-level 99.7/99.9/99.2 on MVTec-AD, 98.9/99.0/96.6 on VisA, 90.5/88.1/81.5 on Real-IAD (I-AUROC/AP/F1-max).
• 4-shot few-shot AD: 97.6 I-AUROC on MVTec-AD, 96.4 on VisA, and 76.7 on Real-IAD, outperforming prior SOTAs by clear margins.
• Zero-shot: trained on Real-IAD and tested on MVTec-AD, achieves 88.0 and 88.7 pixel-level AUROC on MVTec-AD and VisA, outperforming WinCLIP.

Measured effects reported in the paper

Surprising findings

• Only six intrinsic normal prototypes suffice; performance plateaus beyond four and degrades when the number is too large.
• Removing the first residual connection in the decoder helps, since it can directly pass anomalous features into reconstruction.

Limitations

• Logical anomalies closely resembling the background (e.g., misplaced anomalies in MVTec Transistor) can be wrongly extracted as INPs.
• Single-class results on VisA are only competitive rather than best (I-AUROC 98.5 vs Dinomaly 98.9).

Experimental setup

Datasets: MVTec-AD VisA Real-IAD

Benchmarks: MVTec-AD VisA Real-IAD

Metrics: AUROC AP F1-max AUPRO

Baselines: RD4AD UniAD DiAD MambaAD Dinomaly SimpleNet DeSTSeg SPADE PaDiM PatchCore WinCLIP PromptAD

Builds on: RD4AD UniAD Focal Loss PatchCore Dinomaly

Closest papers by meaning

• Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.87)
• ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.87)
• Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.87)
• One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.87)
• AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.86)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.86)