Magnetic Tile Surface Defect Detection Methodology Based on Self-Attention and Self-Supervised Learning.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Magnetic tile defects significantly impact industrial motor quality.
• Existing automatic defect recognition methods struggle with complex defect patterns, leading to low accuracy and difficult practical application.
• Current approaches often involve complicated system structures and ineffective image segmentation and target detection.

Purpose of the Study:

• To develop an improved method for automatic recognition of surface defects on magnetic tiles.
• To enhance the accuracy and practicality of magnetic tile defect detection systems.
• To leverage self-supervised learning (SSL) for more effective feature extraction in defect recognition.

Main Methods:

• A multihead self-attention mechanism was developed for locating and extracting features from defect areas.
• A full-connection neural network was designed for accurate classification of different defect types.
• A knowledge distillation process without labeling simplified the SSL training, utilizing a Vision Transformer (ViT) backbone for feature extraction and a labeled dataset for classification.

Main Results:

• The SSL method extracted richer and more detailed features compared to supervised learning models.
• The integrated system achieved a high testing accuracy of 98.3% on a public magnetic tile surface defect dataset.
• The model demonstrated strong robustness and good generalization ability, handling 5 defect categories and 1 non-defect category.

Conclusions:

• The proposed SSL method effectively addresses the challenges of complex and diverse magnetic tile defect patterns.
• The developed system offers a simpler, more accurate, and robust solution for industrial magnetic tile surface defect recognition.
• This approach significantly improves the quality assessment of magnetic tiles in industrial motors.