Bi-Grid Reconstruction for Image Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised and self-supervised methods excel in general industrial anomaly detection but struggle with fine-grained defects.
• Existing reconstruction-based methods often face challenges like over- or under-detection and the Identical Shortcut (IS) problem.

Purpose of the Study:

• To propose GRAD: Bi-Grid Reconstruction for Image Anomaly Detection, a novel approach for enhanced detection of fine-grained anomalies.
• To improve generalization and detection accuracy in industrial image anomaly detection scenarios.

Main Methods:

• Utilizes two continuous grids as feature repositories to aid reconstruction, enhancing generalization and mitigating the IS problem.
• Introduces an additional grid for abnormal features alongside a normal feature grid to refine normal feature boundaries.
• Incorporates the Feature Block Pasting (FBP) module for synthesizing anomalies at the feature level, enabling rapid deployment of the abnormal grid.

Main Results:

• Demonstrates significant improvement over state-of-the-art methods on classic industrial datasets (MVTecAD, VisA, GoodsAD).
• Achieves enhanced detection performance for fine-grained defects by refining normal feature boundaries.
• Shows suitability for a unified task setting, enabling single-model training for multiple classes.

Conclusions:

• GRAD: Bi-Grid Reconstruction offers a robust solution for fine-grained image anomaly detection in industrial settings.
• The dual-grid approach and FBP module effectively address limitations of existing methods, improving accuracy and generalization.
• The method's adaptability for multi-class detection highlights its practical value in diverse industrial applications.