Axle Surface Defect Detection Based on YOLOv8

In the field of high-speed trains, axles are vital components. Their life and conditions are directly related to the safety of train operations, so it is necessary to carry out regular maintenance on them. This paper presents a method for detecting axle surface defects using the YOLOv8 deep learning algorithm. The method solves the common problem of manual involvement in accurately locating axle surface defects in high-speed trains, and it overcomes traditional algorithms' limitations regarding recognition accuracy and efficiency. The proposed method promotes automation in axle surface defect detection. Through 200 epochs of training with an axle surface defect dataset, the concluding mAP50 (mean Average Precision at Intersection over Union threshold of 0.5) reached 89.94%. Further, the mAP50:95 (mAP at a higher IoU threshold of 0.95) hit 50.37%. The mAP50 improved from 61.78% to 89.94% compared to model validation on the Coco dataset. These results imply that the model effectively detects axle surface defects, indicating its potential to meet rigorous industrial precision detection standards.