Rail surface defect detection based on semantic augmentation and YOLOv8

To unclear defects caused by surface reflection, small sample, high real-time detection requirements, and low accuracy in surface defects, a surface defect instance segmentation detection method combining semantic augmentation and YOLO was proposed. Firstly, using RSDDs dataset to cut out the defect areas and Fourier transform to extract semantic features, combined with the original image to construct surface defect augmentation dataset. Secondly, a detection head that integrates low-level features had been added to the YOLOv8 model, and a detection model for instance segmentation of surface defects was been constructed. Finally, the model was trained and tested, and the performance of semantic augmentation and model improvements were verified through comparative experiments. The experimental results demonstrate that semantic features in the Fourier domain can suppress the influence of surface reflection. Both the semantic augmentation and model improvements can effectively improve the accuracy and recall. The accuracy of detection and instance segmentation are improved by 2.1 and 3.0 on mAP50 by using semantic augmentation. The model improvements on YOLOv8 will bring 1.0 and 1.4 increase on detection and instance segmentation, respectively. By combining semantic augmentation and model improvements, the mAP50 for detection and instance segmentation are achieved to 0.937 and 0.934, respectively, and the mAP50～95 will be improved 11.4 and 11.9 percentage. Besides, proposed method can maintain good real-time performance during significantly improving the accuracy. The research can provide solutions for surface defect detection, and intelligent maintenance management on railway infrastructure. © 2024, Central South University Press. All rights reserved.