作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (9): 288-296. doi: 10.19678/j.issn.1000-3428.0058977

• 开发研究与工程应用 • 上一篇    下一篇

基于改进D-LinkNet模型的高分遥感影像道路提取研究

张立恒1, 王浩1, 薛博维2, 何立明1, 吕悦1   

  1. 1. 长安大学 信息工程学院, 西安 710064;
    2. 西安中科星图空间数据技术有限公司, 西安 710000
  • 收稿日期:2020-07-17 修回日期:2020-09-03 发布日期:2020-09-09
  • 作者简介:张立恒(1995-),男,硕士,主研方向为遥感影像处理、模式识别;王浩、薛博维,硕士;何立明,副教授、博士;吕悦,硕士。
  • 基金资助:
    中央高校基本科研业务费专项资金(300102249302)。

Research of Road Extraction from High-Resolution Remote Sensing Images Based on Improved D-LinkNet Model

ZHANG Liheng1, WANG Hao1, XUE Bowei2, HE Liming1, Lü Yue1   

  1. 1. School of Information Engineering, Chang'an University, Xi'an 710064, China;
    2. Xi'an Zhongkexingtu Space Data Technology Co., Ltd, Xi'an 710000, China
  • Received:2020-07-17 Revised:2020-09-03 Published:2020-09-09

摘要: 针对高分影像中的道路信息易受植被阴影、高楼建筑物、河流等非道路信息干扰的问题,提出一种高分遥感影像道路提取模型。在中心区域引入channel-spatial双注意力机制捕获道路信息的全局特征依赖,并基于原始模型DICE+BCE损失函数,构建新型的超参数权重损失来优化网络模型中参数迭代的误差,改善道路分割的精度。按照1:1、2:1、3:1、4:1、5:1这5种比值设定超参数权重比,通过调节超参数权值比获取模型最佳的道路分割性能。实验结果表明,与FCN-8s、U-Net等模型相比,改进D-LinkNet模型道路分割效果明显提升,能有效地规避因非道路因素对道路提取干扰而导致的“虚检”“漏检”“误检”的现象。

关键词: 高分遥感影像, 双注意力机制, 全局特征依赖, 超参数权重损失, 道路分割

Abstract: The road information in high-resolution images tend to be disturbed by non-road information such as vegetation shadows, tall buildings and rivers.To address the problem, an improved model is proposed to extract road parts from high-resolution remote sensing images.For the construction of the model, a channel-spatial bi-attention mechanism is introduced to capture the global characteristic dependence of road information in the central region.Then the new hyperparameter weight loss is constructed based on the DICE+BCE loss of the original model to reduce the error of the parameter iteration in the network model and improve the accuracy of road segmentation.The hyperparameter weight ratio is successively set to 1:1, 2:1, 3:1, 4:1 and 5:1, and the best road segmentation performance of the model is obtained based on the adjustment of the hyperparameter weight ratio. The experimental results show that compared with FCN-8s, U-Net and other models, the improved D-LinkNet model delivers a significant improvement in road segmentation effect.The algorithm can effectively avoid false detection and missed detection that are caused by interference of non-road factors in road extraction.

Key words: high-resolution remote sensing image, biattention mechanism, global characteristic dependence, hyperparameter weight loss, road segmentation

中图分类号: