摘要
为改善现有深度学习方法获取图像特征尺度单一、提取精度较低等问题,提出多尺度空洞卷积金字塔网络建筑物提取方法。多尺度空洞卷积金字塔网络以U-Net为基础模型,编码-解码阶段采用空洞卷积替换普通卷积扩大感受野,使得每个卷积层输出包含比普通卷积更大范围的特征信息,以利于获取遥感影像中建筑物特征的全局信息,金字塔池化模块结合U-Net跳跃连接结构整合多尺度的特征,以获取高分辨率全局整体信息及低分辨率局部细节信息。在WHU数据集上进行提取实验,交并比达到了91.876%,相比其他语义分割网络交并比提升4.547%~10.826%,在Inria数据集上进行泛化实验,泛化精度高于其他网络。结果表明所提出的空洞卷积金字塔网络提取精度高,泛化能力强,且在不同尺度建筑物提取上具有良好的适应性。
In order to improve the existing deep learning methods for acquiring image feature scales with single scale and low extraction accuracy,building extraction method from multi-scale dilated convolutional pyramid network is proposed based on U-Net.The dilated convolution was used in the en-decoding stage to replace ordinary convolution to expand the receptive field,resulting in the fact that the output of each convolution layer contains a larger range of feature information than ordinary convolution,which facilitates the acquisition of global information about the building features in remote sensing images.The pyramid pooling module is combined with the U-Net jump connection structure to integrate multi-scale features with high-resolution global overall information and low-resolution local detailed information obtained.The extraction experiment was conducted on the WHU dataset,and the intersection ratio reached 91.876%,an increase of 4.547%~10.826%compared to other semantic segmentation networks.The generalization experiment was performed on the Inria dataset,and the generalization accuracy was higher than that by other networks.The results show that the proposed dilated convolutional pyramid network has high extraction accuracy,high generalization ability,and strong adaptability in the extraction of buildings of different scales.
作者
张春森
刘恒恒
葛英伟
史书
张觅
ZHANG Chunsen;LIU Hengheng;GE Yingwei;SHI Shu;ZHANG Mi(College of Geomatics,Xi’an University of Science and Technology,Xi’an 710054,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China)
出处
《西安科技大学学报》
CAS
北大核心
2021年第3期490-497,574,共9页
Journal of Xi’an University of Science and Technology
基金
国家自然科学基金项目(92038301)
陕西省自然科学基金项目(2018JM5103)
中国博士后基金项目(2018M642915)
测绘遥感信息工程国家重点实验室开放基金项目(18R01)。
关键词
建筑物提取
多尺度
空洞卷积
金字塔池化
building extraction
multiscale
dilated convolution
pyramid pooling
