FPN与Mask R-CNN多任务融合的建筑物边界提取  

Multi-task learning for building boundary extraction based on FPN and Mask R-CNN

作　　者：康停军 闫晨曦 张新长 夏义雄 孙颖 KANG Tingjun;YAN Chenxi;ZHANG Xinchang;XIA Yixiong;SUN Ying(Foshan Surveying Mapping and Geoinformation Research Institute,Foshan,Guangdong 528000,China;School of Atmospheric Sciences,Sun Yat-sen University,Zhuhai,Guangdong 519082,China;School of Geography and Remote Sensing,Guangzhou University,Guangzhou 510006,China;School of Geography and Planning,Sun Yat-sen University,Guangzhou 510275,China)

机构地区：[1]佛山市测绘地理信息研究院,广东佛山528000 [2]中山大学大气科学学院,广东珠海519082 [3]广州大学地理科学与遥感学院,广州510006

出　　处：《测绘科学》2023年第6期151-160,180,共11页Science of Surveying and Mapping

基　　金：国家自然科学基金面上项目(42071441);国家自然科学基金青年项目(42201351);广东省基础与应用基础区域联合基金-青年基金项目(2020A1515110441)。

摘　　要：针对目前主流深度学习网络对建筑物边界提取时,存在的边界不准确和内部空洞等问题,该文提出了将语义分割网络-全卷积神经网络(FPN)和实例分割网络-掩膜区域卷积神经网络(Mask R-CNN)进行有机结合的多任务深度学习框架。该方法旨在实现建筑物的语义分割任务和实例分割任务,首先,使用深度残差网络(ResNet50)和全卷积FPN组成的骨干网络(backbone)进行多层次、多尺度的特征提取。然后,将网络分解为两个任务,一个基于FPN的反卷积语义分割任务,另一个则为利用区域候选网络(RPN)的实例分割任务;进一步分析两种结果的特点,通过规则合并两个任务的提取结果。最后,通过建筑物的面积和周长面积比对结果进行后处理,筛选最终的建筑物并进行边界规则化。结果表明,该文方法可以有效利用不同深度学习任务的优势,其精度优于单独使用FPN和Mask R-CNN网络。feature were extracted at multi-level and multiscale.Secondly,the network was split into two branches.The features were sent to two networks for the settlement of the semantic segmentation(FPN)and instant segmentation region proposal networks(RPN).Further analysis was conducted on the characteristics of the two results,and the extraction results of the two tasks were combined by rules at the end of the network.Finally,the combined results were filtered by comparing with the element area and perimeter area, and the final buildings were boundaries wereregularized. The results showed that this multi-task method could benefit from the two different deeplearning tasks to extract building boundaries, and its accuracy was better than using FPN and MaskR-CNN networks alone.

关 键 词：高分辨率遥感影像 建筑物边界提取 多任务 语义分割 实例分割 深度学习 

分 类 号：P237[天文地球—摄影测量与遥感]