Real-Time Railway Track Defect Detection Using an Optimized YOLOv8 Framework

Abstract

Railway track defect detection is vital for ensuring the safety and reliability of train operations and infrastructure. Traditional defect detection methodsboften suffer from missed detections, inaccurate localization, and limited capability in identifying small-scale anomalies, particularly under complex environmental conditions. To overcome these limitations, this work proposes an improved track
defect detection network based on YOLOv8, referred to as DSO-YOLOv8. The proposed model incorporates three major
improvements over the standard YOLOv8 architecture. First, the original detection head is replaced with a decoupled head, which enhances generalization by separately learning object localization and classification features. Second, a small object detection layer is added, extending the feature pyramid structure to better handle multi-scale defect patterns, especially minor cracks and subtle
anomalies. Third, we integrate the Omni- Dimensional Dynamic Convolution (ODConv) into the neck of the model to enable a 4D attention mechanism, which significantly improves the network’s focus on critical defect regions, enhances fine-grained feature extraction, and mitigates the impact of lighting and background clutter. Experimental results demonstrate that the proposed
DSO-YOLOv8 model achieves a mean average precision (mAP) of 98.6%, outperforming the baseline YOLOv8 model by a margin of 3.7%. The enhanced architecture exhibits strong robustness in real-time detection across various defect types and challenging railway scenarios, making it a practical and efficient solution for intelligent railway infrastructure monitoring.