MuSc-V2: Zero-Shot Multimodal Industrial Anomaly Classification and Segmentation With Mutual Scoring of Unlabeled...

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Zero-shot anomaly classification and segmentation aim to detect defects without labeled data.
• Existing methods overlook the distinct property of normal vs. anomalous image patches: normal patches exhibit high similarity, while anomalies are diverse and isolated.

Purpose of the Study:

• To propose a novel Mutual Scoring framework (MuSc-V2) for zero-shot anomaly classification/segmentation.
• To leverage the discriminative property of normal patch similarity versus anomalous patch diversity.
• To support flexible single 2D/3D or multi-modal inputs.

Main Methods:

• Iterative Point Grouping (IPG) for improved 3D representation and reduced false positives.
• Similarity Neighborhood Aggregation with Multi-Degrees (SNAMD) to fuse 2D/3D neighborhood cues for discriminative multi-scale patch features.
• Mutual Scoring Mechanism (MSM) and Cross-modal Anomaly Enhancement (CAE) for score fusion and anomaly recovery.
• Re-scoring with Constrained Neighborhood (RsCon) to suppress false classifications.

Main Results:

• MuSc-V2 achieves significant performance gains: +23.7% AP on MVTec 3D-AD and +19.3% on Eyecandies dataset.
• Outperforms previous zero-shot anomaly detection benchmarks.
• Surpasses the performance of most few-shot anomaly detection methods.
• Demonstrates robust performance on full datasets and smaller subsets, adaptable across diverse product lines.

Conclusions:

• The proposed Mutual Scoring framework effectively utilizes the inherent properties of normal and anomalous patches for superior zero-shot anomaly detection.
• MuSc-V2 offers a flexible and robust solution for industrial anomaly classification and segmentation, advancing the state-of-the-art in unsupervised defect detection.