一种结合稀疏表示和纹理分块的遥感影像超分辨率方法  

Remote Sensing Image Super-Resolution Method Using Sparse Representation and Classified Texture Patches

作　　者：刘帅[1] 朱亚杰[2] 薛磊[2] 

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]北京大学遥感与地理信息系统研究所,北京100871

出　　处：《武汉大学学报（信息科学版）》2015年第5期578-582,共5页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金资助项目(61303128);秦皇岛市科学技术研究与发展计划资助项目(2012023A234)~~

摘　　要：提出了一种基于稀疏表示和纹理分块的单幅遥感影像超分辨率方法,主要利用先验知识及影像自身的纹理信息重构遥感图像。首先,提取用于字典学习的图像块,从高、低分辨率遥感图像块中训练出冗余字典,采用正交匹配追踪方法更新字典,用迭代的方法直到算法收敛;然后,将训练的字典应用于遥感影像超分辨率重构。重构时将图像块分成平滑块和非平滑块两种类型,平滑块采用双三次卷积方法重构,非平滑块采用低分辨率遥感图像块的稀疏表示系数及高分辨率图像块冗余字典重构。实验结果表明,此方法重构速度较快,并在视觉及客观评价指标上有较好的超分辨率效果。A super-resolution method based on sparse representation and classified texture patches was proposed, mainly using the priori knowledge and texture to reconstruct remote sensing images. First, extract image blocks for dictionary learning, the over-complete dictionary was learned from the high and low resolution remote sensing image blocks. Orthogonal match pursuit was used to calculate the sparse conefficients, then the coefficients were fixed, iterative method was used to update the diction- ary until the algorithm converges. Then, the training dictionary was used to reconstruct the remote sensing images. In this step, the image was divided into smooth patches and non-smooth patches, bicubic interpolation was used for smooth patches while sparse conefficients and over-complete diction- ary were used for non-smooth patches. Experiment shows that this method has a faster reconstruction speed and can achieve satisfied super-resolution results in the visual effects and objective evaluation in- dicators.

关 键 词：稀疏表示 超分辨率 字典学习 遥感影像 纹理分块 

分 类 号：P237.3[天文地球—摄影测量与遥感] TP751[天文地球—测绘科学与技术]