摘要
从无人机影像中自动提取建筑物对城乡规划和管理至关重要,然而,在复杂背景干扰和建筑物外观变化很大的情况下给实例提取带来挑战。因此,提出一种改进的Mask区域卷积神经网络(R-CNN)方法用于无人机影像的建筑物自动实例提取。改进方法以ResNet-101作为特征提取网络,在特征融合网络方面,通过添加自底向上的路径增强整个特征层次的定位能力,同时在特征融合中加入空洞空间金字塔池化模块(ASPP)来提高多尺度能力与改善模型性能。在自制建筑物数据集上的综合实验结果表明,与原始的Mask R-CNN方法相比,改进方法的mAP值提高了2.6%,能够很好地实现无人机影像建筑物实例提取。
Automatic building extraction from unmanned aerial vehicle(UAV)images is crucial for urban and rural planning and management.However,it poses challenges to instance extraction in complex background interference and highly variable building appearance.This paper proposed an improved Mask region-based convolutional neural network(R-CNN)method for automatic instance extraction of buildings from UAV images.The improved method used ResNet-101 as the feature extraction network,and the localization ability of the whole feature hierarchy was enhanced by adding bottom-up paths in terms of the feature fusion network.Meanwhile,the atrous spatial pyramid pooling(ASPP)module was added to the feature fusion to increase the multiscale ability and improve the model performance.The comprehensive experimental results on the self-made building dataset show that compared with the original Mask R-CNN method,the mAP value of the improved method is increased by 2.6%,which can well realize the building instance extraction from UAV images.
作者
方超
廖运茂
刘飞
王坚
赵小平
FANG Chao;LIAO Yunmao;LIU Fei;WANG Jian;ZHAO Xiaoping(School of Geomatics and Urban Spatial Information,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;School of Architectural and Survey Mapping Engineering,Beijing Polytechnic College,Beijing 100042,China)
出处
《北京测绘》
2024年第1期97-101,共5页
Beijing Surveying and Mapping
基金
北京建筑大学“双塔计划”项目(JDYC20220825),北京建筑大学青年教师科研能力提升计划(X21021)。
关键词
建筑物提取
Mask
R-CNN
路径融合
空洞空间金字塔池化模块
building extraction
Mask region-based convolutional neural network(R-CNN)
path fusion
atrous spatial pyramid pooling(ASPP)
作者简介
方超(1997-),男,安徽铜陵人,硕士在读,研究方向为遥感图像处理。E-mail:fpermail@163.com;通信作者:刘飞,E-mail:pntrc@cumt.edu.cn。
