西安电子科技大学学报 ›› 2022, Vol. 49 ›› Issue (1): 236-244.doi: 10.19665/j.issn1001-2400.2022.01.025

• 计算机科学与技术&人工智能 • 上一篇    

一种基于边界感知的遥感影像建筑物提取方法

张艳(),王翔宇(),张众维(),孙叶美(),刘树东()   

  1. 天津城建大学 计算机与信息工程学院,天津 300384
  • 收稿日期:2020-12-17 出版日期:2022-02-20 发布日期:2022-04-27
  • 通讯作者: 张众维
  • 作者简介:张 艳(1982—),女,副教授,E-mail: zhangyan@tcu.edu.cn;|王翔宇(1997—),女,天津城建大学硕士研究生,E-mail: 1196245995@qq.com;|孙叶美(1990—),女,助理实验师,E-mail: sunyemei1216@163.com;|刘树东(1965—),男,教授,E-mail: liushudong@tcu.edu.cn
  • 基金资助:
    国家自然科学基金(41971310)

Boundary-aware network for building extraction from remote sensing images

ZHANG Yan(),WANG Xiangyu(),ZHANG Zhongwei(),SUN Yemei(),LIU Shudong()   

  1. School of Computer and Information Engineering,Tianjin Chengjian University,Tianjin 300384,China
  • Received:2020-12-17 Online:2022-02-20 Published:2022-04-27
  • Contact: Zhongwei ZHANG

摘要:

遥感影像的复杂性给建筑物提取研究带来了极大的挑战。深度学习的引入提高了遥感影像建筑物提取的准确率,但仍存在边界模糊、目标漏检和提取区域不完整等问题。针对这些问题,提出了一种基于边界感知的遥感影像建筑物提取网络,该网络包括特征融合网络、特征增强网络和特征细化网络三部分。首先,特征融合网络采用编码-解码结构提取不同尺度特征,并设计了交互聚合模块融合不同尺度的特征;然后,特征增强网络采用减法和级联操作对漏检目标进行学习增强,得到更加全面的特征;最后,特征细化网络使用编码-解码结构对特征增强网络的输出进一步细化,得到丰富的建筑物边界特征。此外,为使网络更加稳定有效,将二值交叉熵损失和结构相似性损失相结合,从像素和图像结构两个层次监督模型的训练学习。通过在数据集WHU上的测试,可知本网络较其他经典算法的客观指标交并比和准确率均有提升,分别达到了96.0%和97.9%;同时主观视觉上提取的建筑物边界更加清晰,区域更加完整分明。

关键词: 建筑物提取, 边界感知, 编码解码, 遥感影像, 深度学习

Abstract:

The complexity of remote sensing images brings a great challenge for building extraction research.The introduction of deep learning improves the accuracy of building extraction from remote sensing images,but there are still some problems such as blurred boundaries,missing targets and incomplete extraction areas.To address these issues,this paper proposes a boundary-aware network for building extraction from remote sensing images,including the feature fusion network,feature enhancement network and feature refinement network.First,the feature fusion network uses the encoding-decoding structure to extract different scale features,and designs the interactive aggregation module to fuse different scale features.Then,the feature enhancement network enhances the learning of missed targets through subtraction and cascade operation to obtain more comprehensive features.Finally,the feature refinement network further refines the output of the feature enhancement network by using the encoding-decoding structure to obtain rich building boundary features.In addition,in order to make the network more stable and effective,this paper combines the binary cross-entropy loss and the structure similarity loss,and supervises the training and learning of the model on both pixel and image structure levels.Through the test on the dataset WHU,in terms of objective metrics,the IoU and Precision of this network are improved compared with other classical algorithms,reaching 96.0% and 97.9% respectively.At the same time,in terms of subjective vision,the extracted building boundary is clearer and the region is more complete.

Key words: building extraction, boundary aware, encoding-decoding, remote sensing image, deep learning

中图分类号: 

  • TP391