YOLOv8DTL: A Deep Transfer Learning Model for Few-Shot Rail Abrasion Detection

Rapid movement of the wheels on defective tracks and high-frequency friction collisions cause vibration between train and rail track, which greatly damages the lifespan of train components and is a significant cause of train derailments. Timely and accurate detection of rail abrasions is of great significance for ensuring the safety of railway operations. Deep learning based-automatic rail abrasion detection methods face the challenge of having fewer samples. Thus, a deep transfer learning framework based on improved YOLOv8 models is developed to detect rail abrasions to address the issue of few-shot learning. Then, deformable convolution networks (DCNs), convolutional block attention module (CBAM) and a new intersection over union (IoU) loss are introduced to improve the detection performance. The ablation experiments show that it effectively reduces the rate of missed and false abrasion detections. By comparing with the existing detection methods, the developed YOLOv8DTL method has higher precision, recall, and average precision under different abrasion size thresholds, indicating that it is more adaptable to detection tasks with different abrasion sizes. It also has the best robustness, maintaining a high level of detection efficiency.