基于压缩感知与结构自相似性的遥感图像超分辨率方法  

Super-resolution Method based on CS and Structural Self-similarity for Remote Sensing Images

作　　者：潘宗序[1] 黄慧娟[1] 禹晶[1] 胡少兴[2] 张爱武[3] 马洪兵[1] 孙卫东[1] 

机构地区：[1]清华大学电子工程系,北京100084 [2]北京航空航天大学机械工程与自动化学院,北京100083 [3]首都师范大学三维空间信息获取与应用教育部重点实验室,北京100037

出　　处：《信号处理》2012年第6期859-872,共14页Journal of Signal Processing

基　　金：国家自然科学基金项目(No.60872083;No.61171117);国家科技支撑计划项目(No.2012BAH31B01)

摘　　要：本文提出了一种基于压缩感知、结构自相似性和字典学习的遥感图像超分辨率方法,其基本思路是建立能够稀疏表示原始高分辨率图像块的字典。实现超分辨率所必需的附加信息来源于遥感图像中广泛存在的自相似结构,该信息可在压缩感知框架下通过字典学习而得到。这里,本文采用K-SVD方法构建字典、并采用OMP方法获取用于稀疏表达的相关系数。与现有基于样本的超分辨率方法的最大不同在于,本文方法仅使用了低分辨率图像及其插值图像,而不需要使用其他高分辨率图像。另外,为了评价方法的效果,本文还引入了一个衡量图像结构自相似性程度的新型指标SSSIM。对比实验结果表明,本文方法具有更好的超分辨率重构效果和运算效率,并且SSSIM指标与超分辨率重构效果具有较强的相关性。A super resolution （SR） method for remote sensing images based on CS, structural self-similarity and diction- ary learning is proposed. The basic idea is to find a dictionary that can represent the high resolution （HR） image sparsely. The extra information comes from the structural self-similarity that widely exist in remote sensing images and this kind of in- formation can be learned through dictionary learning in the CS frame. In this method, we use K-SVD method to find the dic- tionary and OMP method to reveal the sparse representation coefficients. Compared with the traditional sample-based SR methods, the most difference of this method is that we only use the inputted low resolution image and its interpolation image rather than other HR images. In addition, an index called as structural self-similarity （SSSIM） is proposed here to evaluate the extent of structural self-similarity in the image. Results of some comparative experiments show that this proposed method has better SR effect and time efficiency, and the SSSIM index has strong cmTelation with the SR effect.

关 键 词：遥感图像超分辨率 结构自相似性 压缩感知 字典训练 图像质量评价 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]