摘要
目的提出一种基于卷积神经网络的混凝土结构裂缝智能检测方法,实现高效准确的裂缝检测。方法以裂缝图像为分析对象,使用数据扩增和图像分割的方法处理混凝土结构原始裂缝图像,构建裂缝数据集;基于机器视觉理论搭建ResNet18网络模型框架,通过学习率、批量大小、优化函数三个超参数对网络模型进行优化设计,进而使用不同的迁移学习方式训练网络,最后使用混淆矩阵对优化后的模型性能做进一步评判。结果实验结果表明,采用指数衰减式迁移学习方式、批量大小为64、优化函数为M_SGD时,搭建的网络在混凝土裂缝识别任务中准确率最高,达到了97.98%。结论优化后的ResNet18网络提高了裂缝检测精度,具有良好的实用性能,并为网络参数的选择提供借鉴。
In order to achieve efficient and accurate crack detection,this paper proposes an intelligent crack detection method for concrete structures based on convolution neural network.Taking the crack image as the analysis object,the method of data amplification and image segmentation is used to process the original crack image of concrete structure,and the crack data set is constructed;Based on the theory of machine vision,the framework of ResNet18 network model is built.The network model is optimized and designed through three hyperparameters of learning rate,batch size,and optimization function,and then the network is trained using different transfer learning methods.Finally,the performance of optimized model is further evaluated by using confusion matrix.The experimental results show that when the exponential decay transfer learning method is adopted,the batch size is 64,and the optimization function is M_SGD,the network built has the highest accuracy in the task of identifying concrete cracks,reaching 97.98%.The optimized ResNet18 network improves the accuracy of crack detection.It has good practical performance and provides a reference for the selection of network parameters.
作者
孟庆成
万达
吴浩杰
李明健
齐欣
MENG Qingcheng;WAN Da;WU Haojie;LI Mingjian;QI Xin(School of Civil Engineering and Geomatics,Southwest Petroleum University,Chengdu,China,610500;School of Civil Engineering,Southwest Jiaotong University,Chengdu,China,610031)
出处
《沈阳建筑大学学报(自然科学版)》
CAS
CSCD
北大核心
2021年第5期832-840,共9页
Journal of Shenyang Jianzhu University:Natural Science
基金
国家自然科学基金项目(52078442)
四川省教育厅自然科学重点项目(16ZA0058)
四川省科技计划项目(2021YJ0038)
2019教育部产学研协同育人项目(201901273052)。