基于最大似然的超分辨率图像复原方法  

Super-resolution image restoration based on maximum-likelihood estimation

作　　者：汪源源[1] 孙志民[1] 蔡铮[1] 王威琪[1] 

机构地区：[1]复旦大学电子工程系,上海200433

出　　处：《仪器仪表学报》2008年第5期949-953,共5页Chinese Journal of Scientific Instrument

基　　金：国家重点基础研究规划基金(2006CB705707);国家自然科学基金(30570488);上海市重点学科建设(B112)资助项目

摘　　要：从超分辨率图像复原的实际应用出发,提出了一种基于最大似然的序列图像复原方法。在通用线性观测成像模型下,利用改进的奇异值分解法估计成像系统点扩散函数,通过高斯金字塔结构下的光流场方法估计图像间的亚像素级平动,利用基于统计特性的方法估计图像噪声方差等。根据以上参数及部分先验知识,运用最速下降迭代算法实现基于最大似然的超分辨率复原,并对迭代结果进行维纳滤波。计算机模拟实验结果表明,该方法具有较好的复原效果,其性能优于典型的反向迭代投影法。利用实际视频图像的复原实验证明了该方法的实用性。A super-resolution image restoration method based on maximum likelihood （ML） estimation is proposed for the application purpose. With the linear observation imaging model, several imaging parameters were estimated using following methods. The point spread function （PSF） of imaging system was firstly estimated using modified singular value decomposition （SVD） method. The optic flow method based on a pyramid down-sampling structure was then applied to estimate the sub-pixel translations between the noisy low-resolution images. The variance of the imaging noise was estimated using a statistical evaluation method. With these estimated parameters and some prior knowledge, the ML estimation based on the steepest descent （SD） algorithm was applied to reconstruct the superresolution image. The reconstructed result was finally filtered by a 2D Wiener filter. Computer simulation experimental results show that this method can effectively improve the image resolution and have a higher reconstruction performance than typical iterative back projection （IBP） method. Some encouraging results were obtained from the experiments on images sampled by a web-camera, which verifies the practicability of the proposed method.

关 键 词：超分辨率图像复原 线性观测成像模型 奇异值分解 最大似然估计 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]