Surface defect detection method for carbon fiber precursor tow based on transfer learning and YOLOv8

Carbon fiber, owing to its high strength, low density, and exceptional corrosion resistance, has found widespread adoption across numerous industries. As the primary feedstock for its manufacture, carbon fiber precursor tow nevertheless still lacks effective techniques for detecting surface defects. We propose a surface-defect detection method for carbon fiber precursor tow based on transfer learning and YOLOv8. First, optical images of the precursor tow surface are captured using a high-speed line-scan camera and subjected to targeted data-augmentation procedures; then, a YOLOv8-based detection model is trained via transfer learning, resulting in accelerated convergence and improved test-set accuracy. Furthermore, to mitigate the limited dataset size and the class imbalance of defect samples, we applied mixup data augmentation to expand the dataset and balance the defect categories, thereby enhancing the model’s training stability and generalization capability. Test results show that the trained YOLOv8 model for carbon fiber precursor-tow surface defect detection achieves an accuracy of 89.5%, effectively identifying two typical types of surface defects.