基于带有权值混合泛函的盲超分辨率  

Weighted mixed-norm based blind super-resolution algorithm

作　　者：刘刚[1,2] 赵红毅[3] 胡臻龙[4] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所,长春130033 [2]上海电机学院电子信息学院,上海201306 [3]西安理工大学高等技术学院,西安710082 [4]浙江越秀外国语学院,浙江绍兴312008

出　　处：《吉林大学学报（工学版）》2012年第4期1054-1058,共5页Journal of Jilin University:Engineering and Technology Edition

基　　金：'973'国家重点基础研究发展规划项目(2009CB72400102A)

摘　　要：为了克服传统的超分辨率算法的限制(如不能处理带有局部运动图像,因此不适合一般的视频序列;模糊算子被认为是提前知道并且对于每一低分辨率帧都是不变的;超分辨率噪声不是采用高斯分布就是采用拉普拉斯分布的),同时考虑到噪声模型,提出了一个广义的局部权值自适应地混合L1和L2泛函的代价函数。权值会根据配准误差和噪声分布自适应地改变并且惩罚图像中配准错误的部分。本文算法对奇异值具有很强的鲁棒性,同时超分辨率图像和模糊算子可以联合估计出来。主观评价和客观评价同时表明了本文算法的有效性。Conventional Super-Resolution （SR） methods have some limitations. First, most of existing SR algorithms can not cope with local motions and hence not suitable for most video sequences. Second, the blurring operator is assumed to be known in advance and be a constant for all the low- resolution images. Finally, SR noise is assumed to be either Gaussian or Laplacian. To overcome these limitations, a general local cost function is proposed that consists of weighted Ll-and 1.2-norms considering the SR noise model. In this function, the weights are generated according to the error of registration and noise distribution, and the inaccurately registered parts of the image are penalized. Both the super-resolved images and blurring operators are estimated at the same time. Both objective and subjective evaluations demonstrate the effectiveness of the proposed algorithm.

关 键 词：信息处理技术 超分辨率 图像配准 图像融合 

分 类 号：TN911.73[电子电信—通信与信息系统]