Adversarial Shape Learning for Building Extraction in VHR Remote Sensing Images

  • 2021-02-22 18:49:43
  • Lei Ding, Hao Tang, Yahui Liu, Yilei Shi, Lorenzo Bruzzone
  • 2

Abstract

Building extraction in VHR RSIs remains to be a challenging task due toocclusion and boundary ambiguity problems. Although conventional convolutionalneural networks (CNNs) based methods are capable of exploiting local textureand context information, they fail to capture the shape patterns of buildings,which is a necessary constraint in the human recognition. In this context, wepropose an adversarial shape learning network (ASLNet) to model the buildingshape patterns, thus improving the accuracy of building segmentation. In theproposed ASLNet, we introduce the adversarial learning strategy to explicitlymodel the shape constraints, as well as a CNN shape regularizer to strengthenthe embedding of shape features. To assess the geometric accuracy of buildingsegmentation results, we further introduced several object-based assessmentmetrics. Experiments on two open benchmark datasets show that the proposedASLNet improves both the pixel-based accuracy and the object-based measurementsby a large margin. The code is available at: https://github.com/ggsDing/ASLNet

 

Quick Read (beta)

loading the full paper ...