端到端堆叠沙漏网络的遥感影像建筑物轮廓重构

针对如何以端到端可训练的方式重构建筑物轮廓的问题,本文提出了一种端到端多残差模块堆叠沙漏网络的建筑物轮廓重构方法。首先,采用多残差模块堆叠沙漏网络提取建筑物角点和边缘特征;其次,利用角点检测模块匹配对应角点的相对位置,以获取候选角点;然后,通过线段采样将候选角点生成候选轮廓线;最后,线验证模块利用候选线段及多残差模块堆叠沙漏网络得到特征图,并验证每个线段是否为建筑物轮廓线,以获得建筑物轮廓重构结果。试验结果表明,在SpaceNet建筑物数据集上,本文方法能检测出建筑物角点及边缘,并有效实现了以端到端可训练的方式重构建筑物轮廓。

Reconstruction of remote sensing image building contour based on end-to-end stacked hourglass network

Considering the problem that how to reconstruct building contour in an end-to-end trainable way, a method of building contour reconstruction based on end-to-end and multi-residual module stacked hourglass network is proposed. Firstly, the multi-residual module stacked hourglass network is used to extract features of buildings' corners and edges. Secondly, the corners detection module is used to match the relative positions of the corresponding corners to obtain candidate corners, and then the candidate corners are sampled by line sampling module to generate candidate contour lines. Finally, the line verification module used candidate contour lines and feature map obtained by stacked hourglass network verify whether each line segment is a building contour, so as to obtain the result of building contour reconstruction. Experimental results show that on SpaceNet building data set, this method can detect the corners and edges of buildings, and effectively reconstruct the building contour in an end-to-end trainable way.