Adaptive visual detection of industrial product defects

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Visual inspection of industrial product defects is crucial but challenged by limited data and class imbalance.
• Existing deep learning methods struggle with direct application due to these data limitations.
• Transfer learning shows promise but faces cross-dataset bias issues.

Purpose of the Study:

• To develop a universal and adaptive industrial defect detection model.
• To overcome sample scarcity and class imbalance in industrial defect datasets.
• To improve the generalization and performance of visual defect detection.

Main Methods:

• Model-agnostic meta-learning (MAML) for learning across multiple datasets.
• Siamese networks for extracting differential features and minimizing bias.
• Coordinate attention mechanism for region of interest feature enhancement.

Main Results:

• The proposed model demonstrates effectiveness on the new BC defects dataset and public benchmarks.
• Successfully transfers knowledge from known defect datasets to novel anomaly detection tasks.
• Achieves improved defect feature highlighting and detection performance.

Conclusions:

• The developed model offers an effective solution for industrial visual defect inspection with limited data.
• The approach provides a general and adaptive framework for diverse industrial defect detection scenarios.
• The new BC defects dataset and code are released to support further research.