基于ResNet-18网络的桥梁损坏图像分类研究  

A Study on Bridge Damage Image Classification Based on ResNet-18 Network

作　　者：潘宇曜 陈焯辉 林佩欣 陈灵 Pan Yuyao;Chen Zhuohui;Lin Peixin;Chen Ling(Beijing Institute of Technology,Zhuhai,Zhuhai,China;Macao University of Science and Technology,Macao,China)

机构地区：[1]北京理工大学珠海学院,广东珠海 [2]澳门科技大学,中国澳门

出　　处：《科学技术创新》2023年第16期93-96,共4页Scientific and Technological Innovation

基　　金：2022年度广东省科技创新战略专项资金——基于深度学习的聋哑人群手语识别系统的研究(pdjh2022a0706)。

摘　　要：近些年来,传统的桥梁损坏检测方法需要大量的时间以及人力、物力和财力,同时它们具有主观性、难以量化、影响正常交通、周期长、实时性差等缺点和局限性。本研究首先使用桥梁损坏图像作为数据集,然后使用ResNet-18网络对桥梁损坏图像进行分类,并且使用Softmax作为网络输出层的激活函数,使用交叉熵函数作为网络的损失函数。接着进行模型的训练,得出模型在测试集上的准确率为82.99%。最后从模型在测试集上的混淆矩阵与分类报告两个角度,对模型进行评估,得出模型在测试集上平均的F_(1)分数达到83%。In recent years,traditional bridge damage detection methods require a lot of time as well as human,material and financial resources,while they have disadvantages and limitations such as subjectivity,difficult to quantify,affecting normal traffic,long cycle time,and poor real-time.In this research,we first use bridge damage images as the dataset,then use ResNet-18 network to classify bridge damage images,and use Softmax as the activation function of the network output layer and cross-entropy function as the loss function of the network.Then the training of the model was performed,and the accuracy of the model on the test set was obtained as 82.99%.Finally,the model was evaluated in terms of both confusion matrix and classification report on the test set,and it was concluded that the model achieved an average F_(1)-score of 83%on the test set.

关 键 词：桥梁损坏图像 ResNet-18网络 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U446[自动化与计算机技术—计算机科学与技术]