摘要
遥感影像中普遍存在着较强的光照不均匀特性,而当前基于深度学习的建筑物提取方法不能有效应对这一问题。为准确提取复杂背景下高分辨率遥感影像中的典型地物要素,本文提出了一种利用BveNet架构并融合高低级语义信息的地物要素提取算法,在特征提取网络中采用基于图像局部相位的特征检测器,生成对光照变化不敏感的相位一致性特征图,并与残差网络ResNet提取多层级特征进行融合,增强特征的判别能力,提升网络的特征提取效率和算法精度。为了验证本文方法的有效性,在公开数据集上进行了对比实验,实验结果表与未融合光照不敏感特征的原始算法相比,本文方法对光照不均匀条件下的建筑物提取,提升了0.9%的平均精度和2.1%的交并比。
Elumination inhomogeneity is common in remote sensing images,and it cannot be solved with building extraction methods based on deep learning.In order to accurately extract typical features from high-resolution remote sensing images in complex background,a feature extraction algorithm that uses the BiseNet architecture and fuses high-level and low-level semantic information is proposed in this paper.A feature detector based on local phase information is used in feature extraction network te generate a phase consistency feature map that is insensitive to illumination changes,and combined with the multi-level features extracted by residual network ResNet te enhance the feature discriminating ability and improve network feature extraction efficiency and algorithm accuracy.In order to verify the effectiveness of the proposed method,a comparative experiment is carried out using public dataset.The experimental results show that compared with the original algorithm without illumination-insensitive features,this method can improve the average accuracy by 0.9% and the IoU by 2.1% in building extraction under inhomogeneous illumonation conditions.
作者
杨乐
王慧
程挺
徐剑
闫科
YANG Le;WANG Hui;CHENG Ting;XU Jian;YAN Ke(Information Engineering University,Zhengzhou 450001,China;Unit 61287,Chengdu 610000,Chona)
出处
《测绘科学与工程》
2019年第5期43-48,共6页
Geomatics Science and Engineering
