卷积神经网络在桥梁结构表面病害检测中的应用研究  

Detect Diseases on the Surface of Bridge Structures by Convolution Neural Network

作　　者：韩晓健[1] 赵志成 沈泽江 HAN Xiaojian;ZHAO Zhicheng;SHEN Zejiang(Nanjing Tech University, Nanjing210000, China)

机构地区：[1]南京工业大学土木工程学院

出　　处：《结构工程师》2019年第2期106-111,共6页Structural Engineers

摘　　要：计算机视觉检测方法在桥梁结构检测中的使用极大地提高了检测效率,该方法的核心是图像分析处理。研究了深度卷积神经网络在桥梁结构表面病害图像分类识别上的应用。根据桥梁各类病害的统计,将桥梁结构表面病害归纳为裂缝、锈蚀与缺损三大类。通过迁移学习技术,迁移训练AlexNet卷积神经网络,构建了桥梁结构表面病害自动识别模型。对比了5种训练集与验证集的组合,结果表明训练样本的不同组合对模型训练具有一定影响。在不同验证集上,模型的最高正确率为98.21%,模型训练较好。在模型实际应用中,三种病害图像的识别率分别为裂缝86%、锈蚀82%、缺损70%,具有较高的识别正确率,可用于桥梁结构表面病害的快速识别。Computer Vision Inspection System has been applied to detecting the structural diseases of bridge.The detection efficiency is greatly improved by the method.This method relies on image processing.In this paper,we try to use convolution neural network to recognize the diseases on the surface of bridge.Considering the differences between the diseases.The diseases are divided into three categories,such as crack,fault and rust.The model to recognize the diseases is trained by the fine-tuned AlexNet.5 models with different combinations of train and validate sets.The result indicates that different combinations affect the accuracy of the model.The highest accuracy of the model is 98.21%.At the end,the ability to promote is also studied.The accuracy is shown:Crack:86%;Rust:82%;Defect:70%.It is shown that the model has enough ability to be used to recognize the diseases on the surface of the bridge.

关 键 词：钢筋混泥土缺陷 无损检测 图像识别 卷积神经网络 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U445.71[自动化与计算机技术—计算机科学与技术]