A Defect Detection Method for Rail Ultrasonic B-Scan Images Based on a Multi-Scale Dense Attention Network

Abstract

Rail internal defects pose significant threats to railway operational safety, and ultrasonic B-scan imaging provides an effective nondestructive testing approach for visualizing subsurface rail conditions. However, manual interpretation of B-scan images is time-consuming and depends heavily on operator experience. To overcome this problem, a Multi-scale Dense Attention Network (MSDA-Net) is proposed for automatic classification of rail ultrasonic B-scan images. The proposed network integrates dilated convolution, multi-scale dense feature extraction, and attention-based feature recalibration to capture defect features at different spatial scales while enhancing discriminative defect-related regions. A dataset containing 536 ultrasonic B-scan images is constructed for model training and evaluation, and the B-scan images are collected from multiple railway sections. Experimental results show that MSDA-Net achieved an accuracy of 0.909, with a macro Precision of 0.867, macro Recall of 0.845, and macro F1-score of 0.848. Comparative and ablation experiments further demonstrate the effectiveness of the proposed network structure for rail defect classification.