安全与环境学报

针对探地雷达路基病害检测图像人工解译困难的问题，提出一种基于YOLOv5的路基裂缝、疏松病害雷达图像识别方法，并使用1 290张VOC格式标记的雷达图像数据集进行训练和验证。使用卷积注意力模块(Convolutional Block Attention Module, CBAM)、上下文引导模块(Context Guided Block, CG Block)和动态检测头模块(Dynamic Head)对YOLOv5模型进行改进，并进行消融试验以探究每个模块的功能。结果显示，在快速空间金字塔池化层(Spatial Pyramid Pooling Fast, SPPF)前添加CBAM模块可提高模型对关键特征的捕获能力，即在减少参数量的同时实现模型的轻量化。使用Context Guided模块替换YOLOv5模型的neck部分的3个C3模块，可提高局部上下文感知能力，减少模型的参数量和内存占用并提高检测精度。改进的C1-CG-YOLOv5算法的m_AP50达到90.8%,m_AP50-95达到64.2%,参数量减少12.8%,能够有效适用于路基裂缝、疏松病害的识别。

As an advanced non-destructive detection technology and one of the most effective tools available, Ground Penetrating Radar(GPR) has been widely applied in identifying subgrade diseases due to its efficiency and high resolution. However, manual interpretation of GPR images is highly challenging, time-consuming, and susceptible to subjective errors. To address these issues, this study proposes an improved YOLOv5-based method for the automated identification of GPR images depicting subgrade cracks and loosening defects. This method is specifically designed to balance real-time performance with detection accuracy. A carefully constructed dataset comprising 1 290 VOC-format annotated radar images of typical subgrade defects was utilized for model training and validation. Several key architectural enhancements were introduced: the Convolutional Block Attention Module(CBAM) was embedded before the Spatial Pyramid Pooling Fast(SPPF) layer to enhance focus on meaningful features while reducing redundant parameters. Additionally, three C3 modules in the neck of the original YOLOv5 were replaced with Context Guided Blocks(CG Blocks) to strengthen contextual feature perception and improve computational efficiency. Further refinements included a Dynamic Head mechanism to enhance scale-aware representation. Ablation studies demonstrated that the combined improvements from the attention and context-guided modules yielded optimal results. The enhanced C1-CG-YOLOv5 model achieved a mean Average Precision at I_oU threshold of 0.5(m_AP50) of 90.8% and a mean Average Precision at I_oU thresholds of 0.5 to 0.95(m_AP50-95) of 64.2%, significantly outperforming the baseline model. Moreover, the model parameters were reduced by 12.8%, indicating superior lightweight performance. Finally, we developed real-time subgrade disease detection software based on the improved model, validating its detection performance, processing efficiency, and practical utility in GPR image analysis. The proposed approach offers an efficient and reliable tool for automated GPR image interpretation, demonstrating strong potential for practical application in infrastructure maintenance and non-destructive evaluation.