UHA: An Intelligent Uncertainty Map Based Hierarchical Attention Network System for Building Segmentation

Abstract

Satellite images are becoming increasingly high quality with the advancement of high-resolution remote sensing technology. Benefited from the development of deep learning, object segmentation for the high-resolution satellite images has achieved significant improvements in recent years. However, for buildings with multiple scales and almost straight edges in satellite images, the current segmentation methods usually struggle to achieve relatively good results. In this paper, we propose an Uncertainty map based Hierarchical Attention network system (UHA) for building segmentation. UHA aims to capture the information about low-confidence areas of the building's straight edges and improve the segmentation prediction step by step by a hierarchical structure network with three decoder branches. Specifically, we first generate uncertainty maps from the sigmoid prediction maps of the segmentation network. Based on the uncertainty maps, we design an attention module where a transform function transforms the uncertainty maps into attention maps. And then, we introduce the attention module to adjacent branches of the hierarchical network to guide the latter branch by the previous branch, which improves the prediction ability of the following branch smoothly. Finally, we conduct comprehensive experiments that show the proposed system can significantly improve the performance upon baselines by a large margin.