Real-Time Analysis of Track Surface Defect Images Using WT-YOLO12

Abstract

The rapid and precise defect detection of rail surface defects is critical for railway operational safety. The conventional inspection methods, however, are often labor-intensive, resource-consuming, and vulnerable to noise. To address these challenges, this article proposes a real-time defect segmentation and detection model called wavelet transform (WT)-YOLO12. First, this article introduces the multiscale convolutional attention (MSCA) mechanism to capture long-range contextual information, compensating for the limited receptive field of the YOLO12 backbone network. Second, this article enhances high-frequency defect features using WTs, significantly alleviating the bottleneck in fine-grained track surface defect segmentation with YOLO12. Extensive experiments on the Types I and II track surface defect datasets validate the effectiveness of the proposed segmentation algorithm. On the Type II railway surface defect datasets (RSDDs) dataset, WT-YOLO12 improves mask mean average precision (mAP) by 11% over the baseline YOLO12s, while maintaining a processing speed of 71.2 frames per second (FPS), which satisfies real-time application requirements. Furthermore, the comparative experiments on the railway dataset demonstrate the model’s strong generalization, yielding a 3.9% increase in mAP over the baseline YOLO12s.