基于语义分割的隧道渗漏水轻量化检测模型研究  

Research on lightweight detection model of tunnel water leakage based on semantic segmentation

作　　者：王丹丹[1] 侯公羽[1] 张欣怡 陈钦煌 邵耀华 符欢欢 WANG Dandan;HOU Gongyu;ZHANG Xinyi;CHEN Qinhuang;SHAO Yaohua;FU Huanhuan(School of Mechanics and Civil Engineering,China University of Mining&Technology,Beijing 100083,China)

机构地区：[1]中国矿业大学(北京)力学与土木工程学院,北京100083

出　　处：《铁道科学与工程学报》2024年第12期5264-5275,共12页Journal of Railway Science and Engineering

基　　金：国家自然科学基金委员会与2020年度高速铁路基础研究联合基金资助项目(U2034205)。

摘　　要：针对传统计算机视觉渗漏水检测算法模型参量大、检测速度慢和复杂背景下抗干扰能力差的问题,提出一种基于DeepLabV3+改进的轻量化分割模型SC-DeepLabV3+,以实现地铁隧道渗漏水病害的高效检测。首先,采用融合空间和通道激励注意力机制的(concurrent spatial and channel squeeze and excitation,简称scSE)轻量化网络S-Efficientnet作为主干特征提取网络,以降低模型参数量,提高渗漏水检测速度,其中,scSE注意力机制通过在空间和通道2个维度上进行信息的压缩和增强,以突出有用通道,抑制无用通道,进而提升模型的检测精度。其次,设计一种高层语义特征细化模块C-ASPP(Contact-Atrous Spatial Pyramid Pooling)。该模块通过缩小膨胀因子、调整特征提取分支数以及融合深度可分离卷积,在考虑全局上下文信息的同时,增强对渗漏水边缘细节的关注;且C-ASPP模块可通过多尺度感受野特征提取结构,更准确地处理和整合来自不同尺度的语义特征,以实现对多尺度语义信息的相互交融,进而增强复杂环境下模型的抗干扰能力。最后,在构建的渗漏水混合数据集上进行实验。研究结果表明:S-Efficientnet主干网络极大地降低模型的复杂度,增强了对渗漏水区域的关注度;C-ASPP模块提高了对多尺度信息的综合利用;SC-DeepLabV3+模型交并比达到90.17%,模型大小仅为5.457 M,图像处理速度可达89.525 f/s。与当今主流语义分割模型相比,SC-DeepLabV3+模型在检测精度和分割速度等方面展现出显著的优越性。To solve the problems of large model parameters,slow detection speed,and poor anti-interference ability in complex backgrounds of traditional computer vision water leakage detection algorithms,a lightweight segmentation model SC-DeepLabV3+based on DeepLabV3+was proposed to realize the efficient detection of tunnel water leakage.First,the lightweight network S-Efficientnet,which integrates spatial and channel squeeze and excitation,was used as the backbone feature network.It can reduce model parameters and improve the speed of water leakage detection.The scSE attention mechanism compresses and enhances information in both spatial and channel dimensions to highlight the useful channels and suppress the useless channels,which in turn improves the detection accuracy of the model.Second,a high-level semantic feature refinement module Contact-Atrous Spatial Pyramid Pooling was designed.This module enhances the attention to the details of water leakage edges while considering the global contextual information by narrowing the expansion factor,adjusting the number of feature extraction branches,and fusing the depth separable convolution.The C-ASPP module processed and integrated semantic features from different scales to realize the mutual intermixing of the multi-scale semantic information and enhanced the anti-interference ability in complex environment.Finally,experiments were conducted on the constructed water leakage hybrid dataset.The results show that the S-Efficientnet backbone network greatly reduces the complexity of the model and enhances the focus on the water seepage area;the C-ASPP module improves the comprehensive utilization of multi-scale information;the intersection and merger ratio of the SC-DeepLabV3+model reaches 90.17%,the size of the model is only 5.457 M,and the image processing speed is up to 89.525 f/s.Compared with existing mainstream semantic segmentation models,the SC-DeepLabV3+model demonstrates significant superiority in terms of detection accuracy and segmentation speed.

关 键 词：深度学习 语义分割 隧道病害 渗漏水 DeepLabV3+ 轻量化 

分 类 号：U45[建筑科学—桥梁与隧道工程]