Acta Geodaetica et Cartographica Sinica ›› 2024, Vol. 53 ›› Issue (7): 1444-1457.doi: 10.11947/j.AGCS.2024.20230056

• Cartography and Geoinformation • Previous Articles     Next Articles

Two-stream boundary constraints and relativistic generation adversarial network for building contour regularization

Jichong YIN(), Fang WU(), Renjian ZHAI, Yue QIU, Xianyong GONG, Ruixing XING   

  1. Institute of Geospatial Information, Information Engineering University, Zhengzhou 450001, China
  • Received:2023-03-20 Published:2024-08-12
  • Contact: Fang WU E-mail:jichongy@whu.edu.cn;wufang_630@126.com
  • About author:YIN Jichong (1997—), male, PhD candidate, majors in intelligent processing of geospatial data. E-mail: jichongy@whu.edu.cn
  • Supported by:
    The National Natural Science Foundation of China(42201491);The Natural Science Foundation for Distinguished Young Scholars of Henan Province(212300410014)

Abstract:

Building extraction from high-resolution remote sensing images is still a hot and difficult research topic in the field of remote sensing application and cartography. Although the introduction of deep learning method has greatly improved the accuracy of building segmentation, the problems of irregular contour and unclear boundary in building segmentation mask still exist. In order to obtain regular building contours and clear boundaries, this paper proposes a method of building contour regularization based on two-stream boundary constraint and relativistic generation adversarial network. The network is composed of a two-stream boundary constraint generator and a relativistic average discriminator. The two-stream boundary constraint generator fuses the boundary details of remote sensing images and input labels through the two-stream network architecture and boundary loss function, thus generating regular building contours. The relativistic average discriminator forces the generator to generate a more realistic building mask by evaluating the quality difference between the ground truth label and the generated regularization result. To verify the performance of the model and explore the reasons for the performance improvement, two groups of experiments were designed on WHU building dataset and Inria aerial image labeling dataset, including a comparative experiment and ablation experiment. The experimental results show that this method can generate regularization results that match ground truth labels, and it has obvious advantages in solving the problems of blurred boundaries and irregular contours of segmentation masks.

Key words: building contour regularization, boundary constraint, generative adversarial networks, high-resolution remote sensing image

CLC Number: