结合迁移学习和Inception-v3模型的路面干湿状态识别方法  

Method of dry and wet state recognition of pavement based on migration learning and Inception-v3 model

作　　者：杨炜[1] 周凯霞 刘佳俊 张志威 王童 YANG Wei;ZHOU Kaixia;LIU Jiajun;ZHANG Zhiwei;WANG Tong(School of Automobile,Chang’an University,Xi’an 710064,China)

机构地区：[1]长安大学汽车学院

出　　处：《中国科技论文》2019年第8期912-916,共5页China Sciencepaper

基　　金：中央高校基本科研业务费专项资金资助项目(300102229112);陕西省自然科学基础研究计划项目(2018JQ5213)

摘　　要：针对传统方法对路面干湿状态识别分类正确率较低的情况,提出了基于迁移学习的路面干湿状态识别方法。利用深度卷积神经网络强大的特征学习和表达能力,自动学习干湿路面的特征,并采用迁移学习的方法将Inception-v3模型在ImageNet图像数据集上学习得到的知识深度迁移至路面干湿状态识别任务。实验结果表明,所提算法在测试集上测得的分类准确率约为94.5%,与非迁移学习算法和基于底层视觉特征识别学习的算法相比,具有更高的准确性和良好的鲁棒性,以及较强的泛化能力。In view of the fact that traditional methods had a low accuracy in classification of pavement wet and dry state,a new recognition method based on migration learning was proposed.Using the powerful feature learning and expression ability of deep convolution neural network,the features of wet and dry pavement were automatically learned.The Inception-v3 model acquired from ImageNet image data set was migrated to the task of identifying wet and dry pavement state by the method of migration learning.The experimental results show that the classification accuracy of the proposed algorithm is about 94.5%on the test set.Compared with the non-transfer learning algorithm and the low-level visual feature recognition learning algorithm,the proposed algorithm has higher accuracy,better robustness and stronger generalization ability.

关 键 词：道路工程 干湿状态识别 深度学习 迁移模型 Inception-v3模型 

分 类 号：U461.91[机械工程—车辆工程]