基于纹理增强卷积网络的居民区要素提取  

Residential area extraction based on texture information enhanced CNN

作　　者：刘松林[1,2] 高凯 张丽[1,2] 李毅恒 LIU Songlin;GAO Kai;ZHANG Li;LI Yiheng(Xi'an Research Institute of Surveying and Mapping,Xi'an 710054,China;State Key of Geo-Information Engineering,Xi'an 710054,China)

机构地区：[1]西安测绘研究所,陕西西安710054 [2]地理信息工程国家重点实验室,陕西西安710054

出　　处：《测绘科学与工程》2019年第5期60-63,72,共5页Geomatics Science and Engineering

摘　　要：本文针对空间分辨率为2m的光学卫星影像中的居民区要素,利用纹理信息对U-Net网络进行了改进,提出了一种基于纹理增强CNN的居民区要素提取方法。首先,使用灰度级量化方法提取影像的纹理信息,并对其进行归一化;然后,在U-Net网络的1×1卷积层之前融合CNN特征和影像纹理信息,并使用融合后的特征继续前向传播计算损失;最后,通过损失反向传播实现网络训练。本文的创新点在于将影像纹理信息融入CNN特征,提高了语义分割结果的精度。通过在天绘一号卫星数据集上进行的对比实验结果表明,本文方法能够获得较高精度的居民区要素提取结果。Aiming at the residential elements in optical satellite images with spatial resolution of 2 meters,the U-Net network is improved by using texture information,and a method of extracting residential elements based on texture enhanced convolutional neurat network(CNN)is proposed.Firstly,the texture information of the image is extracted using the gray level quantification method and then it will be normalized by maximum value.Secondly,the CNN features and texture information image are meraed before the 1×1 convelution layer of the U-Net,and the fused features are used to continue foeard propagation to calculate the loss.Finally,the network training is realized through the back propagation of loss.The innovation point of this paper is to integrate image texture information into CNN features and improve the accuracy of semantic segmentation results..Experimental results based on TH-1 satellite images show that the proposed alyorithm is more accurate than existing algorithms.

关 键 词：卷积神经网络 纹理特征 光学影像 语义分割 地物提取 

分 类 号：P237[天文地球—摄影测量与遥感]