• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于预测和残差细化网络的道路提取算法研究

熊炜1,2,管来福1,王传胜1,童磊1,李利荣1,刘敏1   

  1. (1.湖北工业大学电气与电子工程学院,湖北 武汉 430068;
    2.美国南卡罗来纳大学计算机科学与工程系,南卡 哥伦比亚 29201)
  • 收稿日期:2019-09-09 修回日期:2019-11-30 出版日期:2020-04-25 发布日期:2020-04-25
  • 基金资助:

    国家留学基金(201808420418);国家自然科学基金(61571182,61601177);湖北省自然科学基金(2019CFB530)

Road extraction algorithm based on
prediction and residual refinement networks
#br#  

XIONG Wei1,2,GUAN Lai-fu1,WANG Chuan-sheng1,TONG Lei1,LI Li-rong1,LIU Min1   

  1. (1.School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;
    2.Department of Computer Science and Engineering,University of South Carolina,Columbia,SC 29201,USA)
  • Received:2019-09-09 Revised:2019-11-30 Online:2020-04-25 Published:2020-04-25

摘要:

针对航拍图像中的道路检测问题,提出了一种基于预测和残差细化网络的航拍图像道路提取算法。首先,预测网络进行初始预测,为了提高分割网络的细化能力,学习到更高层的道路特征信息,预测网络中引入了空洞卷积和多核池化模块。其次,残差细化网络对预测网络的输出进一步细化,改善预测网络结果出现的模糊问题。此外,针对航拍图像中道路像素比例较小的特点,网络还融合了二元交叉熵、结构相似性以及交并比损失函数,以减少道路信息损失。在Massachusetts道路数据集上的实验结果表明,精确率、召回率、F值和准确率等指标分别达到了99.3%,95.7%,97.3%和95.1%,交并比及平均结构相似性评价指标也分别达到了94.8%和84.3%,相比于其他算法,该算法有一定的应用价值。
 
 

关键词: 航拍图像, 道路提取, 深度学习, 空洞卷积, 多核池化, 损失函数

Abstract:

Aiming at the problem of road detection in aerial images, a road extraction algorithm based on prediction and residual refinement networks is proposed. Firstly, the prediction network makes initial predictions. In order to improve the refinement ability of the segmentation network and learn higher- level road features, the dilated convolution and multi-kernel pooling modules are introduced in the prediction networks. Secondly, the residual refinement network will further refine the output of the prediction network and improve the ambiguity of the prediction network results. In addition, considering the small proportion of road pixels in aerial images, the network also combines binary cross entropy, structural similarity, and intersection over union loss functions to reduce road information loss. The experimental results on the Massachusetts road dataset show that the precision, recall, F value and accuracy reaches 99.3%, 95.7%, 97.3% and 95.1%, respectively. The intersection over union and structural similarity also reaches 94.8% and 84.3%, respectively. Compared with other algorithms, this proposed algorithm has certain application value.

 

 

 

Key words: aerial image, road extraction, deep learning, dilated convolution, multi-kernel pooling, loss function