多尺度特征融合深度学习建筑物的提取方法  

Multi-scale Feature Fusion Deep Learning-based Building Extraction Method

作　　者：刘恒恒 张春森[1] 葛英伟 史书 LIU Hengheng;ZHANG Chunsen;GE Yingwei;SHI Shu(College of Geomatics,Xi'an University of Science and Technology,Xi'an 710054,China)

机构地区：[1]西安科技大学测绘科学与技术学院,陕西西安710054

出　　处：《地理空间信息》2022年第2期97-100,共4页Geospatial Information

基　　金：国家自然科学基金重大研究计划(9203830001);陕西省自然科学基金资助项目(2018JM5103)。

摘　　要：提出一种基于多尺度特征融合的建筑物提取方法,结合新的网络DenseASPP-UNet,以实现影像多尺度特征的融合,进而高精度提取建筑物。通过Inria开源建筑物航空影像数据集进行验证,表明DenseASPP-UNet相比其他深度学习方法建筑物提取精度有很大的提升。We proposed a building extraction method based on multi-scale feature fusion.Based on the UNet model,we jumped layers in parallel with the dense cavity space pyramid pooling layer and used 1×1 convolution to reduce feature mapdimension.We proposed a new network Den-seASPPUNet to achieve multi-scale imagefeature fusion and extracted buildings with high precision.We verified it with the Inriaopen source building aerial image data set.The experiment result shows that DenseASPP-UNetgreatly improves building extraction accuracy compared with other deep learning methods.

关 键 词：深度学习 建筑物提取 多尺度特征融合 密集空洞空间金字塔池化 

分 类 号：P237[天文地球—摄影测量与遥感]