Deep learning-based weak micro-defect detection on an optical lens surface with micro vision.

Area of Science:

• Computer Vision
• Materials Science
• Manufacturing Engineering

Background:

• Manual quality control in optical lens (OL) production suffers from inefficiency and unreliability.
• Existing automated methods struggle with detecting weak, micro-sized defects due to low resolution and ambiguous morphology.

Purpose of the Study:

• To develop an automatic micro vision-based inspection system (MVIS) for optical lens surface defect detection.
• To propose a novel deep learning algorithm, ISE-YOLO, for improved detection of weak micro-defects.

Main Methods:

• An MVIS system was developed to capture defect images and create an OL dataset for predictive inference.
• A deep learning algorithm, ISE-YOLO, was designed incorporating an ISE attention mechanism and a novel class loss function.
• The model was optimized for deep layers to extract richer semantics and learn more information from convolution layers.

Main Results:

• ISE-YOLO demonstrated superior performance on the OL dataset compared to YOLOv5, with significant increases in mean average precision (3.62%), recall (6.12%), and F1 score (3.07%).
• Compared to YOLOv7, ISE-YOLO achieved a 2.58% higher mean average precision.
• ISE-YOLO also reduced weight size by over 30% and increased detection speed by 16% relative to YOLOv7.

Conclusions:

• The proposed ISE-YOLO algorithm effectively addresses the challenges of detecting weak micro-defects in optical lens production.
• The MVIS system integrated with ISE-YOLO offers a more efficient and reliable solution for automated quality control in optical lens manufacturing.