为解决现有目标检测模型在隧道衬砌渗漏水检测中，因模型适配性不强而难以准确检测复杂背景下尺寸变化剧烈、边界模糊及受噪声干扰病害的问题，提出YOLOv11-CMDS模型，针对主干网络中C3K2模块对多尺度特征输入处理不足的问题，通过引进多尺度MSCB模块并与C3K2模块融合，创建C3K2-MSCB模块，以加强模型主干网络对尺寸变化剧烈病害的特征提取能力。此外，为解决原上采样在处理图像特征时引起的图像边缘细节失真，引入动态上采样DySample，自适应地使采样点集中在渗漏水病害关键细节，增强模型对病害边界模糊的细节还原能力。最后，嵌入SENetV2注意力机制，通过全局信息压缩和通道权重重分配方式，进一步增强模型对病害的感知能力。试验以自主采集的隧道衬砌渗漏水数据集为样本，结果表明： 1）消融试验中，YOLOv11-CMDS模型的mAP50%、Precision、Recall、F1分数分别为84.8%、81.7%、74.7%、78.0%，相较于基线模型YOLOv11n分别提升了3.1%、3.8%、2.2%、3.0%，且在对比试验中，该模型整体性能优于现有经典目标检测模型。2）该模型通过多尺度特征增强、动态细节还原与注意力机制的协同作用，有效提升了隧道衬砌渗漏水病害的检测精度，为隧道工程的智能化管理与维护提供了相关的技术支撑。
The existing target detection models are difficult to accurately detect diseases with severe size change, blurred boundaries, and noise interference in complex backgrounds due to the weak adaptability in tunnel lining leakage detection. To address this challenge, a YOLOv11-CMDS model is proposed. For insufficient processing of multiscale feature inputs by the C3K2 module in the backbone network, a C3K2-MSCB module is created by introducing a multiscale MSCB module and integrating it with the original C3K2 module to enhance the feature extraction ability of the model’s backbone network for diseases with severe size changes. In addition, to address the distortion of image edge details caused by the original upsampling in image feature processing, DySample is introduced to adaptively concentrate the sampling points on the key details of water leakage diseases, and enhance the model’s ability to restore the details of blurred disease boundaries. Finally, the SENetV2 attention mechanism is embedded to further enhance the model’s ability to perceive diseases through global information compression and channel weight redistribution. Experiments are conducted on a self-collected tunnel lining leakage dataset as the test sample. The results show that: (1) In the ablation experiment, the mAP50%, Precision, Recall, and F1 scores of the YOLOv11-CMDS model are 84.8%, 81.7%, 74.7%, and 78.0%, respectively, which are 3.1%, 3.8%, 2.2%, and 3.0% higher than those of the baseline model YOLOv11n, respectively. (2) Through synergistic effect of multiscale feature enhancement, dynamic detail restoration, and the attention mechanism, the model effectively improves the detection accuracy of tunnel lining water leakage diseases.
