融合边界与语义特征的遥感影像语义分割方法  

Semantic Segmentation of Remote Sensing Image Based on Fusion of Boundary and Semantic Features

作　　者：徐进军 吴玉炜 徐卓琳 王晗 XU Jinjun;WU Yuwei;XU Zhuolin;WANG Han(Jiangxi Institute of Land Space Survey and Planning,Nanchang 330025,China;School of Land Resource and Environment,Jiangxi Agricultural University,Nanchang 330025,China)

机构地区：[1]江西省国土空间调查规划研究院,江西南昌330025 [2]江西农业大学国土资源与环境学院,江西南昌330045

出　　处：《江西测绘》2024年第4期5-8,共4页JIANGXI CEHUI

摘　　要：针对遥感图像语义信息复杂多样以及地物边界不清晰导致地物分割精度不高的问题,论文提出了一种融合边界与语义特征网络模型(BFNet),该模型通过语义流和边界流同时在编码阶段提取语义和边界特征信息,在解码阶段集成同尺度语义和边界特征,从而提高语义分割的性能。与语义分割方法相比,在WHDLD公开数据集上,BFNet模型的可视化结果的地物更清晰,OA、mIoU和F1分数等三个评估指标比U-Net高2.5%、3%、3.4%。Aiming at the problem of low accuracy in ground object segmentation due to the complex and diverse semantic information in remote sensing images,coupled with unclear boundaries of ground objects,this paper proposes a Boundary and Semantic Feature Network model(BFNet).This model extracts semantic and boundary feature information simultaneously during the encoding stage through semantic and boundary streams,and integrates semantic and boundary features at the same scale during the decoding stage,thereby enhancing the performance of semantic segmentation.Compared with semantic segmentation methods,on the WHDLD public data set,the ground objects of the visualization results of BFNet model are clearer,and the three evaluation indicators such as OA,mIoU and F1 scores are 2.5%,3%and 3.4%higher than those of UNet.

关 键 词：遥感影像语义分割 边界特征 语义特征 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] P237[自动化与计算机技术—控制科学与工程]