Research on Rail Defect Recognition Method Based on Improved RT-DETR Model

This is the authors' abstract. We don't add key points for this paper.

An advanced rail defect recognition method is presented in this paper, which utilizes an improved Real-Time DEtection Transformer (RT-DETR) model to increase the precision of detecting internal rail defects and reduce false alarms. Traditional manual and deep learning methods have been struggling with small target detection and background noise interference. To address these challenges, several modifications are proposed, including integrating the Bottle2Neck module in the backbone network, adopting the RepBi-PAN structure to enhance multi-scale feature extraction, employing Wise-IoU (WIoU) to improve small target localization, and a hard negative sample optimization strategy. Experimental results based on the rail defect dataset show that our method achieves 97.2% on mAP and 2.54% on false alarm rate, which proves the effectiveness of our method in rail defect detection.