PSgANet: Polar Sequence-Guided Attention Network for Edge-Related Defect Classification in Contact Lenses

Area of Science:

• Industrial process optimization
• Artificial intelligence in manufacturing
• Contact lens quality control

Background:

• Contact lens manufacturing demands high precision and safety.
• Existing AI methods struggle with subtle defects in the rim-connected zone (RCZ).
• Need for advanced AI for enhanced operational efficiency and quality assurance.

Purpose of the Study:

• Introduce a novel AI approach for accurate contact lens defect detection.
• Develop a customized deep learning model, PSgANet, for RCZ defect identification.
• Improve accuracy and classification of defects in contact lens manufacturing.

Main Methods:

• Employed polar coordinate transformation to linearize edge irregularities.
• Developed the Polar Sequence-guided Attention Network (PSgANet) integrating sequence learning and attention mechanisms.
• Compared PSgANet (GRU, LSTM, Transformer) with conventional models (GoogleNetv4, EfficientNet, Vision Transformer).

Main Results:

• PSgANet models outperformed traditional Convolutional Neural Network (CNN)-based models.
• LSTM-based PSgANet achieved the highest accuracy, precision, and recall.
• Demonstrated up to a 7.75% accuracy improvement over GoogleNetv4.

Conclusions:

• The proposed PSgANet method is effective for detecting and classifying RCZ defects.
• AI integration, specifically PSgANet, enhances quality control in contact lens production.
• This approach offers a promising solution for stringent safety requirements in medical device manufacturing.