基于CNN的公路路网自动提取方法研究  

Research on Automatic Extraction Method of Highway Network Based on CNN

作　　者：昌宏哲 杜红静 袁洋 CHANG Hong-zhe;DU Hong-jing;YUAN Yang(Highway Administration Bureau of Henan Provincial Transportation Department,Zhengzhou,Henan 450018,China;Henan Transportation Development Research Institute Co.,Ltd.,Zhengzhou,Henan 450018,China)

机构地区：[1]河南省交通运输厅公路管理局,河南郑州450018 [2]河南交通发展研究院有限公司,河南郑州450018

出　　处：《计算技术与自动化》2022年第2期164-167,共4页Computing Technology and Automation

摘　　要：为提高深度学习网络中公路提取性能,提出了一种基于CNN框架的公路路网自动提取方法。为进一步提高模型训练效果,设计了一种考虑公路结构的损失函数。计算损失时不仅考虑每个像素的重要程度,同时考虑公路的全局结构。通过仿真分析,所提方法准确率达到92.4%,较传统CNN方法提高13%。In order to improve the performance of road network extraction in deep learning network,an automatic extraction method of road network based on CNN framework is proposed.In order to further improve the training effect of the model,a loss function considering the highway structure is designed too.When calculating the loss,not only the importance of each pixel is considered,but also the global structure of highway is considered.Through simulation analysis,the accuracy of the proposed method reaches 92.4%,which is 13%higher than that of the traditional CNN method.

关 键 词：CNN 公路路网 自动提取 交叉熵损失 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]