Multi-defect detection and classification for aluminum alloys with enhanced YOLOv8

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Aluminum alloy defects impact structural integrity and safety.
• Current defect detection methods lack accuracy and precise localization.

Purpose of the Study:

• To propose an enhanced YOLOv8-ALGP model for improved aluminum alloy defect detection and classification.
• To address limitations in accuracy and target framing in existing methods.

Main Methods:

• Dataset augmentation with 3 additional defects for improved generalization.
• Integration of an ALGC3 module combining Ghost convolution and residual connections for model lightweighting.
• Reconstruction of the backbone network with enhanced neck layers for improved feature extraction.

Main Results:

• Reduced model parameters from over 300,000 to 160,000.
• Increased average detection accuracy from 64.5% to 71.3% compared to YOLOv8.
• Significantly improved detection accuracy for specific defects like effacement (21.6% to 32.2%) and jet (38.5% to 60%).

Conclusions:

• The proposed YOLOv8-ALGP method effectively detects and classifies aluminum alloy surface defects.
• The enhanced model offers improved accuracy and efficiency, contributing to better product performance in the aluminum industry.