Accurate Industrial Anomaly Detection and Localization Using Weakly-Supervised Residual Transformers.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Industrial anomaly detection (AD) benefits from training with anomalous samples.
• Extensive pixel-level annotations are often impractical for real-world AD applications.
• Existing methods require significant manual labeling efforts.

Purpose of the Study:

• Introduce a novel framework, WeakREST, for accurate AD with minimal manual annotations.
• Reduce dependency on precise pixel-level labels in anomaly detection.
• Achieve high anomaly detection and localization accuracy using weak supervision.

Main Methods:

• Reformulate pixel-wise localization into block-wise classification.
• Introduce Positional Fast Anomaly Residuals (PosFAR) for effective anomaly feature representation.
• Adapt Swin Transformer and develop ResMixMatch for weak label learning.

Main Results:

• WeakREST achieves 83.0% AP on MVTec-AD in unsupervised settings, surpassing prior work.
• Attains 87.6% AP in supervised settings, outperforming previous bests.
• Achieves 87.1% AP using only bounding box annotations, exceeding pixel-supervised methods.

Conclusions:

• WeakREST demonstrates superior performance in anomaly detection and localization across multiple datasets.
• The framework effectively leverages weak annotations, significantly reducing labeling costs.
• WeakREST sets a new benchmark for weakly-supervised industrial anomaly detection.