首页|基于DC-Unet的混凝土桥梁表观裂缝识别方法

基于DC-Unet的混凝土桥梁表观裂缝识别方法

扫码查看
为解决服役混凝土桥梁表观损伤检测存在误检率高、受背景噪声影响大等问题,提出一种基于DC-Unet的表观裂缝识别方法.首先,使用密集残差连接模块(DRCM)替换U-net模型中每次下采样和上采样前以及最后1×1卷积前的卷积操作,增加了模型深度;其次,在模型第1次上采样前的位置融入空洞空间池化金字塔模块,扩大了模型感受野,提升了模型获取多层次表观裂缝特征的能力;最后,将结合了空间和通道的注意力模块以残差连接的方式融入DRCM结构中,提高了模型对表观裂缝特征区域的关注,使用Labelme软件标注50张3 648像素×2 736像素分辨率的表观裂缝图像,基于窗口滑动算法构建了包含1 347张混凝土表观裂缝图像及标签图像的TimCracks数据集,将提出的表观裂缝识别方法与U-net模型、SegNet模型、U-net++模型、传统OTSU阈值分割算法和基于Canny算子的边缘检测算法进行了比较.结果表明:提出的方法能准确分割识别混凝土桥梁表观裂缝,具有精度高和抗噪性好等优势,在表观裂缝图像背景存在涂层干扰及不光滑褶皱状况下均可有效降低误检率,裂缝识别准确率、交并比和F1-score分别达96.28%、73.80%和84.91%,3个评价指标较U-net模型、SegNet模型、U-net++模型均有提升,与传统裂缝分割算法相比,提出的DC-Unet网络解决了传统方法的误检问题,能将裂缝从涂层背景中有效分割.
Surface crack identification method of concrete bridge based on DC-Unet
To solve the problem of largely false detection rate and highly background noise for the surface damage detection of existing concrete bridges,a surface crack damage identification method was proposed based on DC-Unet.Firstly,the dense residual connection module(DRCM)was used to replace the convolution operation before each downsampling,and upsampling and the last 1X1 convolution in the U-net model,which increased the model depth.Secondly,the atrous spatial pyramid pooling module was integrated into the position before the first upsampling of the model,which expanded the receptive field of the model and improved the ability of the model to obtain multi-level apparent fracture characteristics.Finally,the convolutional block attention module combined with space and channel was integrated into the DRCM structure in a residual connection way,which improved the model's attention to the apparent crack feature area.The Labelme software was used to label 50 apparent crack images with a resolution of 3 648 pixel X 2 736 pixel,and a TimCracks dataset containing 1 347 concrete apparent crack images and label images was constructed based on the window sliding algorithm.The proposed concrete bridge apparent crack recognition method was compared with U-net model,SegNet model,U-net++model,traditional OTSU threshold segmentation algorithm and edge detection algorithm based on Canny operator.The results show that the proposed method can accurately segment and identify the apparent cracks of concrete bridges,and has the advantages of high precision and good noise resistance.It can effectively reduce the false detection rate under the condition of coating interference and unsmooth folds in the background of apparent crack images.The accuracy,intersection over union and F1-score of crack recognition are 96.28%,73.80%and 84.91%,respectively.The three evaluation indexes are improved,compared with U-net model,SegNet model and U-net++model.Compared with the traditional crack segmentation algorithm,the proposed DC-Unet network solves the problem of false detection of traditional methods.The cracks can be effectively segmented from the coating background.3 tabs,13 figs,28 refs.

bridge engineeringconcrete crackU-net structuredense residual connectionatrous spatial pyramid poolingdamage detection

马亚飞、孙文康、何羽、王磊

展开 >

长沙理工大学土木工程学院,湖南长沙 410114

桥梁工程 混凝土裂缝 U-net结构 密集残差连接 空洞空间池化金字塔 损伤检测

国家重点研发计划项目湖南省自然科学基金创新研究群体项目

2021YFB26009002020JJ10060

2024

长安大学学报(自然科学版)
长安大学

长安大学学报(自然科学版)

CSTPCD北大核心
影响因子:1.011
ISSN:1671-8879
年,卷(期):2024.44(3)
  • 28