A Multi-Category Anomaly Editing Network With Correlation Exploration and Voxel-Level Attention for Unsupervised...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Surface anomaly detection is challenging due to product category variations.
• Existing feature editing methods for anomaly detection are limited to specific product types.
• Auto-encoders often over-generalize, hindering accurate anomaly reconstruction.

Purpose of the Study:

• To develop a unified, multi-category surface anomaly detection model.
• To overcome limitations of existing methods in generalizing across product categories.
• To improve the accuracy and precision of anomaly segmentation.

Main Methods:

• A multi-category anomaly editing network with a dual-branch training approach (normal and anomaly branches).
• Utilized multi-category anomaly feature editing based auto-encoder (MCAFE-AE) for reconstruction and inpainting.
• Incorporated dual-entropy constrained deep embedded clustering (DEC-DECM) and patch-based adaptive thresholding (PAT) in the normal branch.
• Employed multi-category anomaly feature editing module (MCAFEM) in the anomaly branch.
• Developed correlation exploration and voxel-level attention based prediction network (CEVA-Net) with correlation-dependency exploration and voxel-level attention refinement module (CDE-VARM) for segmentation.

Main Results:

• The proposed method achieves state-of-the-art performance on three benchmark datasets.
• Demonstrated superior performance in detecting and segmenting surface anomalies across multiple product categories.
• The dual-branch approach effectively handles variations in normal and anomalous features.

Conclusions:

• The developed multi-category anomaly editing network provides a robust solution for surface anomaly detection.
• The novel architecture and training strategy significantly improve generalization capabilities.
• This work advances the field of anomaly detection in industrial inspection.