With the accelerated advancement of rail transit systems, conventional manual inspection methodologies exhibit significant limitations in both efficiency and accuracy. Despite the advantages of deep learning-based methods in automated defect detection, they still face challenges such as defect diversity, interference from complex environments, and computational resource constraints. This study proposes a visual inspection framework for rail defects based on an enhanced YOLOv7-tiny model, which integrates the Convolutional Block Attention Module (CBAM) for channel-spatial dual-dimensional attention mechanisms and Deformable Convolutional Networks (DCN) for depth wise separable convolutions, thereby improving the recognition capability of small-sized defects under complex environmental conditions. A novel lightweight hybrid parallel network architecture is proposed, incorporating depth wise separable convolution and channel pruning methodologies to enhance computational efficiency. Experimental results demonstrate that the improved model achieves a 95.7% Mean Average Precision (mAP) on the railway defect dataset, representing a 7.3% enhancement over the baseline YOLOv7-tiny. The optimized network exhibits a 42% reduction in parameters, decreasing from 6.0M to 3.5M, and a 53% reduction in computational complexity, dropping from 13.2 GFLOPs to 6.2 GFLOPs. These improvements highlight its significant advantages for deployment on resource-constrained detection devices. The study presents an innovative solution for intelligent track maintenance, achieving an optimal balance between precision and operational efficiency.
Du et al. (Sat,) studied this question.