摘要
混凝土裂隙几何信息识别的精确度,影响后期工程的安全。而传统的检测方法存在对裂隙识别不准、不全、不即时的缺陷,无法满足精度和实效性的现实需求。本文提出一种融合自注意力机制与全卷积神经网络的图像分割算法,以混凝土裂隙图像建立数据集,搭建深度学习网络;以全卷积神经网络训练模型,使用空间自注意力模块调整特征编码,输出基于自注意力机制模块识别的高精度二值图。经精准率、召回率、平均交并比和综合评价指标等维度同传统图像分割方法进行对比,结果显示,本文方法得到的混凝土裂隙二值图与原图最相近,在定量上精准率、召回率、平均交并比和综合评价指标分别达到62.93%,88.08%,72.21%和83.86%,进而验证本文提出的方法优于传统方法裂隙识别方法。
The accuracy of the geometric information identification of concrete cracks will affect the safety of later projects.However,the traditional detection methods have the defects such as inaccurate,incomplete,and not instantaneous identification of cracks,which cannot meet the practical requirements of accuracy and effectiveness.This paper proposed an image segmentation algorithm that combines a self-attention mechanism with a fully convolutional neural network.A concrete crack was used to build a data set to construct a deep learning architecture;a fully convolutional neural network was used to train the model and a spatial self-attention module was used to adjust the characteristic encoding,output high-precision binary figure based on self-attention mechanism module recognition.The dimensions of precision rate,recall rate,average merge ratio and comprehensive evaluation index were compared with traditional image segmentation methods.The results show that the binary figure of concrete cracks obtained by this method is the closest to the original image.In terms of quantitative accuracy,recall rate,average crossover ratio and comprehensive evaluation index reached 62.93%,88.08%,72.21%and 83.86%,respectively,and then verified that the method proposed is superior to the traditional method of crack identification.
作者
哈纳提·吐尔森哈力
林杭
Hanat tursenhali;LIN Hang(School of Resources&Safety Engineering,Central South University,Changsha 410083,China)
出处
《铁道科学与工程学报》
CAS
CSCD
北大核心
2021年第4期844-852,共9页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(51774322)
湖南省自然科学基金资助项目(2018JJ2500)。
关键词
深度学习
全卷积神经网络
自注意力机制
裂隙识别
deep learning
fully convolutional networks
self-attention
crack identification
作者简介
通信作者:林杭(1980−),男,福建福州人,教授,博士,从事数值计算与岩土力学等方面研究,E−mail:linhangabc@126.com。
