Efficient surface defect identification for optical components via multi-scale mixed Kernels and structural...

Area of Science:

• Optics and Materials Science
• Computer Vision and Machine Learning

Background:

• Optical surface defect identification is crucial for quality control in manufacturing.
• Challenges include defect subtlety, complexity, and scale variation.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for identifying various optical surface defects.
• To address the limitations of existing methods in handling diverse defect characteristics.

Main Methods:

• A deep network employing multi-scale mixed kernels to capture features at different receptive fields.
• An asymmetric mixed kernel design for rotationally robust feature extraction.
• Structural re-parameterization to optimize the model for fast inference.

Main Results:

• The proposed method achieved 97.39% accuracy on an optical surface defect dataset.
• Inference speed reached 201.76 frames/second with only 5.23M parameters.
• Demonstrated superior performance in identifying both manufacturing and non-manufacturing defects.

Conclusions:

• The developed deep network offers an effective solution for optical surface defect identification.
• The model's balance of accuracy and speed meets practical industrial requirements.
• Structural re-parameterization enhances deployability without compromising performance.