Building Extraction Based on Semantic Segmentation Networks with Different Skeletons
Using high-resolution remote sensing images for building extraction may result in missing and incorrect extraction of building edge lines.Replacing encoder convolutional layers with skeletons can solve these problems to some extent.This paper uses three different skeletons to improve DeeplabV3+ and UNet Deep Learning networks.Using the WHU and Inria datasets for verification,the results show that the improved network with the introduction of three skeletons improves accuracy by 0.49%,1.52%,and 0.87%compared to DeeplabV3+ on the WHU dataset,respectively.The accuracy of UNet network improves by 1.15%,3.24%,and 3.13%,respectively.The same conclusion can be drawn on the Inria data,which solves the problems of missing and miss extraction of edge lines to some extent.
国家科技期刊平台
NSTL
万方数据
