Related Experiment Videos
Region-Prompt-Guided Anomaly Detection With Entropy-Based Consistency Modeling
Summary
This study introduces RPGAD, a novel framework for multi-class visual industrial anomaly detection. RPGAD effectively addresses cross-class confusion and pixel similarity issues, improving detection accuracy in complex scenarios.
Area of Science:
- Computer Vision
- Machine Learning
- Industrial Automation
Background:
- Visual industrial anomaly detection has advanced to complex multi-class settings.
- Existing methods struggle with cross-class semantic confusion and pixel similarity failures.
- These limitations hinder accurate anomaly detection in diverse industrial applications.
Purpose of the Study:
- To propose RPGAD (Region-Prompt Guided Anomaly Detection), an information-theoretic framework for robust multi-class anomaly detection.
- To model anomalies as semantic predictive instability using dual-path responses.
- To overcome limitations of current approaches in handling complex visual patterns and inter-class similarities.
Main Methods:
- Developed DPENet (Dual-Path regional Energy evaluation Network) for region prompt generation via entropy-guided energy.
- Integrated RDNet (Reverse Distillation Network) for selective reconstruction of prompted regions.
- Employed Prototype-Contrastive Optimal Transport (PCOT) loss for enhanced inter-class separability and feature aggregation.
Main Results:
- RPGAD demonstrated strong performance across five benchmarks (MVTecAD, VisA, BTAD, MPDD, Real-IAD).
- Achieved high mean Anomaly Detection (mAD) scores on all tested datasets.
- Showcased significant pixel-level Average Precision (AP) and F1-max gains over baseline methods.
Conclusions:
- RPGAD provides accurate and robust multi-class anomaly detection and localization.
- The proposed framework effectively handles complex visual scenarios and semantic confusions.
- RPGAD represents a significant advancement in industrial anomaly detection technology.
