基于卷积神经网络的建筑物精细化提取  

Refine Extraction of Buildings Based on the Convolutional Neural Network

作　　者：黄小赛[1,2] 李艳 马佩坤[1,2] 高扬 吴剑亮 

机构地区：[1]南京大学国际地球系统科学研究所,江苏南京210023 [2]江苏省地理信息技术重点实验室,江苏南京210023

出　　处：《地理空间信息》2018年第3期97-100,共4页Geospatial Information

基　　金：国家自然科学基金资助项目(41371331)

摘　　要：现有图像分割方法往往受图像模糊和噪声的影响,提取的轮廓不准确。为了提取建筑物的精确轮廓,提出了一种基于卷积神经网络的集成方法,包括建筑物定位、形状判断、形状匹配等步骤。实验证明,无论是对DSM图像还是多光谱影像,该方法都能获得精确的建筑物轮廓。The existing methods of image segmentation are often affected by image blurring and noise, so that the extracted contours are not accurate. In order to extract the accurate contours of buildings, this paper proposed an integrated approach based on the convolutional neural network, which included building locating, shape judgment and shape matching. The experimental result shows that the proposed method can obtain accurate building contours both for DSM and multispectral images.

关 键 词：建筑物提取 卷积神经网络 先验形状 

分 类 号：P237[天文地球—摄影测量与遥感]