一种利用卷积神经网络重建的遥感影像Curvelet融合方法  

A Curvelet Fusion Method of Remote Sensing Image by Convolutional Neural Network Reconstruction

作　　者：李彦东 隋立春[1,2] 陈楠 袁欢欢 徐家利 LI Yandong;SUI Lichun;CHEN Nan;YUAN Huanhuan;XU Jiali(College of Geological Engineering and Geomatics,Chang’an University,Xi’an 710054,China;National Administration of Surveying,Mapping and Geoinformation Engineering Research Center of Geographic National Conditions Monitoring,Xi’an 710054,China)

机构地区：[1]长安大学地质工程与测绘学院,西安710054 [2]地理国情监测国家测绘地理信息局工程技术研究中心,西安710054

出　　处：《遥感信息》2022年第3期93-100,共8页Remote Sensing Information

基　　金：陕西交通运输厅科研项目(18-06K、16-01K)。

摘　　要：针对传统多光谱与全色影像融合方法容易产生畸变、忽略多光谱影像本身的空间细节特征等问题,提出了一种基于超分辨率卷积神经网络与Curvelet变换的影像融合方法,以提升多光谱影像的空间细节,加强其与全色影像的相关性,减少融合产生的畸变。该方法首先利用高分辨率全色影像进行超分辨率重建学习,利用学习得到的网络参数对多光谱影像进行超分辨率卷积神经网络重建,提升其空间细节特征;其次,在Gram-Schmid变换融合基础上,根据Curvelet变换具有保持影像空间细节的特点,将全色影像与替换分量进行融合;最后,通过逆变换得到高分辨率遥感影像。实验结果表明,该算法在影像光谱信息和空间细节表达能力上,整体优于其他传统算法,且对不同数据具有很好的适应性。A novel fusion algorithm to Multispectral(MS)and Panchromatic(PAN)remote sensing image based on super-resolution convolutional neural network and Curvelet transform,which can enhance the spatial details of the MS,improve the correlation with PAN and reduce the distortion,is proposed for the traditional fusion methods can cause image distortion and ignored the spatial detail characteristics of MS.Firstly,using high-resolution PAN images to obtain the super-resolution reconstruction convolutional neural network model,MS images are reconstructed by the model to enhance spatial details.Then,because Curvelet transform has the characteristic of keeping image space detail,in the Gram-Schmidt transform,PAN image is fused with the first component of GS through the Curvelet transform.Finally,high resolution remote sensing image is synthesized by GS inverse transformation.The experimental results demonstrate that the proposed method can attain better image spectral information and spatial detail simultaneously compared with other methods.Meanwhile,the proposed method has a certain adaptability to different data.

关 键 词：空谱融合 超分辨率影像重建 卷积神经网络 Curvelet变换分析 Gram-Schmid变换 

分 类 号：P237[天文地球—摄影测量与遥感]