一种基于改进YOLOv8的重载铁路轨道缺陷检测方法

A Method for Detecting Defects on Heavy-Load Railway Tracks Based on the Improved YOLOv8 Model

摘要: 针对重载铁路轨道场景中背景纹理复杂(如道床碎石干扰)、缺陷目标细小(如螺栓、扣件)以及裂纹特征难以提取导致检测精度不足的问题，文章提出了一种基于改进YOLOv8的重载铁路轨道缺陷检测方法。首先，利用ConvNeXt V2网络替换原有的主干，通过大核卷积与GRN层设计，显著增强了模型的多尺度特征表达与长程依赖建模能力，有效抑制了复杂背景干扰。其次，引入SFS-Conv (空频选择卷积)替代部分标准卷积，利用空间–频率双域协同建模与通道适应选择机制，强化了对钢轨裂纹高频纹理及细小扣件结构的感知敏感度。实验结果表明，改进后的模型在包含3870张图像的轨道缺陷数据集上表现优异，mAP@0.5达到95.0%，相比原始YOLOv8在精度、召回率及综合检测性能上均有显著提升。该方法有效解决了复杂环境下螺栓缺失与钢轨裂纹的误检、漏检问题，具有较高的工程应用价值。

Abstract: In heavy-haul railway track scenarios, complex background textures (e.g., ballast gravel interference), tiny defect targets (e.g., bolts and fasteners), and the difficulty of extracting crack features often result in insufficient detection accuracy. To address these challenges, this paper proposes an improved YOLOv8-based method for heavy-haul railway track defect detection. The original backbone is replaced with ConvNeXt V2, which leverages large-kernel convolutions and Global Response Normalization (GRN) layers to significantly enhance multi-scale feature representation and long-range dependency modeling, effectively suppressing complex background interference. Additionally, Space-Frequency Selection Convolution (SFS-Conv) is introduced to replace selected standard convolutions, enabling joint spatial-frequency domain modeling and adaptive channel selection, thereby markedly improving sensitivity to high-frequency crack textures and fine-grained fastener structures. Experimental results on a track defect dataset comprising 3870 images demonstrate that the improved model achieves an mAP@0.5 of 95.0%, with substantial gains in precision, recall, and overall performance over the baseline YOLOv8. The proposed approach effectively mitigates missed and false detections of bolt loss and rail cracks in complex environments, exhibiting strong practical engineering value for heavy-haul railway maintenance.