Defect Detection Method of Carbon Fiber Unidirectional Band Prepreg Based on Enhanced YOLOv8s.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Existing carbon fiber prepreg surface defect detection faces challenges with small targets and elongated defects.
• Inaccuracies in current methods hinder precise quality control for carbon fiber materials.

Purpose of the Study:

• To develop enhanced YOLOv8s-based models for improved carbon fiber prepreg surface defect detection.
• To address limitations in detecting small and large aspect ratio defects.

Main Methods:

• Integration of Global Attention Mechanism (GAM) and Deformable Large Kernel Attention (DLKA) into YOLOv8s architecture.
• Attention mechanisms inserted between backbone and detection head for enhanced feature extraction.
• Evaluation of YOLOv8s-GAM and YOLOv8s-DLKA models for defect identification.

Main Results:

• YOLOv8s-GAM achieved 84.4% mAP@0.5, 79.3% precision, and 78.3% recall.
• YOLOv8s-DLKA achieved 86.4% mAP@0.5, 82.4% precision, and 80.2% recall.
• Both models showed significant improvements over the original YOLOv8s, particularly in precision and recall for defect detection.

Conclusions:

• Enhanced YOLOv8s models with GAM and DLKA offer effective solutions for precise defect identification on carbon fiber prepreg surfaces.
• The proposed models improve detection accuracy for challenging defect types, enhancing quality assurance.
• These advancements contribute to more reliable manufacturing processes for carbon fiber composites.