江西科学2025，Vol.43Issue(5)：914-919,6.DOI:10.13990/j.issn1001-3679.2025.05.016
基于多尺度空洞卷积FCN网络的建筑物提取
Building Extraction Based on Multiscale Atrous Convolutional FCN Network
摘要
Abstract
To address issues of contextual information loss and partial pixel omission when ex-tracting building using Fully Convolutional Networks(FCNs),this study introduces multi-scale atrous convolution into the FCN framework and proposes the Multiscale Atrous Convo-lution FCN8s(MAC-FCN8s)deep learning network.By replacing standard convolutions with atrous convolutions of varying scales,the receptive field is expanded,enabling im-proved contextual feature capture.Additionally,an optimal loss function is employed to pro-mote more effective model learning.In this study,the Massachusetts dataset was used for experiments.The method proposed in this paper was compared with the results of building extraction with FCN32s,FCN16s,and FCN8s for building extraction.A generalization ex-periment was also performed on the Satellite Dataset I(Global cities)dataset.Results show that the MAC-FCN8s network achieves higher building extraction accuracy,improves ex-traction completeness,and reduces instances of partial pixel loss.Moreover,the method ex-hibits good generalization performance on the Global cities dataset.关键词
多尺度空洞卷积/FCN/损失函数/建筑物提取Key words
multiscale atrous convolution/FCN/loss function/building extraction分类
天文与地球科学引用本文复制引用
李星,张宝金,贾鲁净,郭子睿,孟硕林..基于多尺度空洞卷积FCN网络的建筑物提取[J].江西科学,2025,43(5):914-919,6.基金项目
国家自然科学基金项目(41571346). （41571346）
