摘要
为充分发挥Transformer模型在高分辨率桥梁裂缝图像分割上的优势,提出了一种基于Transformer和坐标注意力(Coordinate Attention,CA)机制的精细化级联分割方法(级联CATransUNet)。首先,开发了一个基于TransUNet的裂缝特征提取模块,用于从低分辨率裂缝图像中初步提取3个尺度的粗粒度裂缝特征;其中,CA机制被引入TransUNet的跳跃连接结构,从而增强TransUNet网络对微小裂缝特征的表征。然后,基于所提取的三尺度粗粒度裂缝特征设计了2个基于物理级联结构的精细化运算模块,实现了从全局和局部2个维度依次恢复裂缝主体与边缘区域的细粒度像素特征。此外,为了充分利用多尺度特征在裂缝边界的细粒度特征表征中的优势,在训练过程中引入了一个带有主动边界回归项的多尺度级联损失。在基于无人机所采集的桥梁高分辨率裂缝图像上开展的消融性试验证明了所提出各组件的有效性。最后,在4 K分辨率的桥梁裂缝图像上开展了对比试验,结果表明:级联CATransUNet在不增加显卡内存需求的前提下,相较于此前最先进的基于传统卷积神经网络(Convolutional Neural Network,CNN)搭建的高分辨率图像精细化网络Segfix和CascadePSP,平均交并比(mean Intersection over Union,mIoU)和平均边缘精度(mean Boundary Accuracy,mBA)分别提升了5%和7%以上。采用级联CATransUNet可实现对高分辨率裂缝图像的精细化分割,为检测人员提供更加全面、准确的结构裂缝信息,从而为结构安全状况评估以及维护决策制定提供技术支撑。
To fully leverage the advantages of the Transformer model in high-resolution(HR)bridge crack image segmentation,a refined cascaded segmentation method,Cascade CATransUNet,based on the Transformer and Coordinate Attention(CA)mechanism was proposed.Firstly,a TransUNet-based crack feature extraction module was introduced to preliminarily extract coarse-grained crack features at three scales from low-resolution(LR)crack images.The CA mechanism was incorporated into the skip-connection structure of TransUNet to enhance the representation of subtle crack features.Then,based on the extracted coarse-grained crack features at the three scales,two refined operation modules based on physical cascaded structures were designed to sequentially restore fine-grained pixel features of the crack body and edge region from both global and local dimensions.Additionally,to fully utilize the advantages of multi-scale features in the fine-grained representation of crack boundaries,a multi-scale cascaded loss with an active boundary regression term is introduced during the training process.Ablation experiments conducted on HR bridge crack images captured by the unmanned aerial vehicle(UAV)demonstrated the effectiveness of each proposed component.Finally,the comparative experiment conducted on 4 K-resolution bridge crack images revealed that the Cascade CATransUNet surpasses the state-of-the-art high-resolution(HR)refinement networks Segfix and CascadePSP,both of which rely on traditional convolutional neural networks(CNNs).Notably,the Cascade CATransUNet achieved significant enhancements of 5.04%and 7.10%in mean Intersection over Union(mIoU)and mean Boundary Accuracy(mBA),respectively,while retaining identical GPU memory requirements.By adopting the Cascade CATransUNet,it becomes feasible to perform fine-grained segmentation of HR crack images,enabling structural inspectors to obtain more comprehensive and accurate crack information.Consequently,this provides valuable technical support for bridge safety assessment and maintenance decision-making processes.
作者
褚鸿鹄
袁华青
龙砺芝
邓露
CHU Hong-hu;YUAN Hua-qing;LONG Li-zhi;DENG Lu(College of Civil Engineering,Hunan University,Changsha 410082,Hunan,China;The Bartlett Faculty of the Built Environment,University College London,London WCIE 6BT,UK;Hunan Provincial Key Laboratory for Damage Diagnosis of Engineering Structures,Changsha 410082,Hunan,China)
出处
《中国公路学报》
EI
CAS
CSCD
北大核心
2024年第2期65-76,共12页
China Journal of Highway and Transport
基金
国家自然科学基金项目(52278177)
湖南省科技创新领军人才项目(2021RC4025)
湖南省研究生科研创新项目(QL20210106)
国家留学基金项目(202206130068)。
作者简介
褚鸿鹄(1995-),男,浙江湖州人,工学博士研究生,E-mail:chuhonghu@hnu.edu.cn;通讯作者:邓露(1984-),男,湖南双峰人,教授,博士研究生导师,工学博士,E-mail:denglu@hnu.edu.cn。
