改进SegNet+CRF高分辨率遥感影像建筑物提取方法  

Building Extraction from High-resolution Remote Sensing Images Based on Improved SegNet and CRF Method

作　　者：赵昊罡 崔红霞[1] 张芳菲 顾海燕 穆潇莹 ZHAO Haogang;CUI Hongxia;ZHANG Fangfei;GU Haiyan;MU Xiaoying(College of Information Science and Technology,Bohai University,Jinzhou 121003,China;CLP Taiji(Group)Co.,Ltd.,Beijing 100083,China;Chinese Academy of Surveying and Mapping,Beijing 100086,China)

机构地区：[1]渤海大学信息科学技术学院,辽宁锦州121003 [2]中电太极(集团)有限公司,北京100083 [3]中国测绘科学研究院,北京100086

出　　处：《计算机测量与控制》2023年第7期177-183,共7页Computer Measurement &Control

基　　金：自然资源部测绘科学与地球空间信息技术重点实验室项目(2020-2-4);辽宁省教育厅重点攻关项目(LZ2020004)。

摘　　要：将传统的语义分割SegNet网络用于高分辨率遥感影像的建筑物提取时,分割的建筑物存在边界模糊、精度较低、错检漏检等问题;为了解决上述问题,提出一种改进SegNet网络+CRF语义分割方法;编码阶段的最低分辨率层引入空洞金字塔池化模型(ASPP,atrous spatial pyramid pooling),通过并行的空洞卷积操作扩大特征提取的感受野;解码阶段构建特征金字塔(FPN,feature pyramid networks)实现特征多尺度融合,弥补上采样过程中丢失的特征信息;最后,预测图像送入全连接条件随机场模型(CRF,fully connected/dense CRF)进行后处理,优化提取的建筑物边缘;实验表明,相较于原SegNet网络,改进方法的建筑物提取像素精度、召回率、平均交并比分别提高了0.48%、1.29%、2.36%。When traditional semantic segmentation SegNet networks are used to extract buildings from high-resolution remote sensing images,there are the problems of fuzzy boundary,low accuracy,error detection and missing detection.In order to solve above problems,an improved SegNet and CRF semantic segmentation method is proposed.An Atrous Spatial Pyramid Pooling(ASPP)model is introduced to the lowest resolution layer in the encoding stage,parallel multiple hole convolutions are used to extract the feature maps of different receptive fields of an image with the expansion rates.In the decoding stage,Feature Pyramid Networks(FPNs)are built to realize multi-scale feature fusion,and compensate the lost feature information during sampling.Finally,based on the fully connected conditional random field(CRF)mode,the prediction images are reprocessed to optimize the building edges.Experimental results show that the building extraction accuracy of the improved model is higher than that of the original SegNet network,the average pixel accuracy,recall and average cross-over ratio are improved by 0.48%,1.29%and 2.36%respectively.The improved method is feasible,and can be extended to similar networks.

关 键 词：语义分割 空洞金字塔池化模型 特征金字塔 全连接条件随机场 迁移学习 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]