【摘要】 高速公路路基病害检测的准确性与高效性对于道路养护至关重要。然而,现有的检测网络常局限于卷积结构，对于多尺度的特征提取与跨尺度结构空间的特征融合还有待加强。笔者提出了一种针对上述问题的CA＿SD＿YOLO网络模型。该模型在骨干网络引入卷积注意力融合模块（CAFM），通过双支路提取特征并融合，卷积支路提取局部细节，注意力支路建模全局关联，实现局部卷积特征与全局注意力特征的互补提取；在特征融合层提出空间扰动感知模块（SDPM）替换原concat模块，以高层特征为参考加权低层特征，实现跨尺度像素级对齐解决跨尺度特征语义与结构信息的错位。本实验结果表明，与YOLOv13相比，在GPR＿Data数据集上整体mAP50、召回率和精确度分别提升了5.6%、5.6%和2.9%。该模型有效提升了复杂场景下的检测能力，为道路养护部门提供高精度、低延迟的自动化病害检测工具，为GPR公路病害自动识别提供了可行的解决方案。更多还原

【Abstract】 The accuracy and efficiency of highway subgrade defect detection are of vital importance for road maintenance. However,the existing detection networks are often limited to convolutional structures, and there is still room for improvement in multi-scale feature extraction and cross-scale spatial feature fusion. This paper proposes a CA＿SD＿YOLO network model to address these issues.This model introduces a convolutional attention fusion module(CAFM) in the backbone network. It extracts features through two branches and fuses them. The convolutional branch extracts local details, while the attention branch models global correlations,achieving complementary extraction of local convolutional features and global attention features. In the feature fusion layer, a spatial perturbation-aware module(SDPM) is proposed to replace the original concat module, using high-level features to weight low-level features, thereby achieving pixel-level cross-scale alignment and resolving misalignment between cross-scale semantic and structural information. The experimental results show that compared with YOLOv13, on the GPR＿Data dataset, the overall mAP50, recall rate,and accuracy have increased by 5.6%, 5.6%, and 2.9%, respectively. This model effectively enhances detection capability in complex scenarios, providing high-precision, low-latency automated defect-detection tools for road maintenance departments and a feasible solution for automatic identification of GPR road defects.更多还原