基于稀疏正则优化图像复原算法.docVIP

基于稀疏正则优化图像复原算法文章编号:1001-9081(2012)01-0261-03 doi:10.3724/SP.J.1087.2012.00261 摘 要:为提高图像复原的速度，改进图像复原的质量，提出一种新算法。将图像复原表示为一类标准的优化问题，采用交替最小化把该优化问题分解为等价的两个子问题。通过迭代求解这两个子问题，获得图像复原问题的解。在此迭代过程中，引入迭代软阈值法处理图像降噪子问题。实验对不同类型的模糊图像进行了复原，其结果验证了算法的有效性。与多级阈值Landweber(MLTL)算法和快速收缩阈值算法(FISTA)相比，处理相同图像时，所提算法可分别节省28%和71%的时间，同时复原图像的信噪比(SNR)可提高0.7~3.5?dB ?关键词: 图像复原；约束优化问题；稀疏表示；交替最小化；迭代软阈值 ?中图分类号: TP391.413 文献标志码:A ? Abstract: For speeding up image restoration and improving the restored results, a new algorithm was proposed. The image restoration was represented as a class of standard optimization problem, which was decomposed into two subproblems by the alternating minimization algorithm. By iteratively solving the two subproblems, a solution to the image restoration problems was obtained. During the subproblem solving, the iterative soft-thresholding algorithm was introduced for the denoising subproblem. In the experiment, the images blurred by different type of blur were restored. The experimental results show the effectiveness of the proposed algorithm. When dealing with the images, compared with Multilevel Thresholded Landweber (MLTL) and Fast Iterative Shrinkage-Thresholding Algorithm (FISTA), the proposed algorithm can reduce the time by 28% and 71% respectively, and it improves the Signal-to-Noise Ratio (SNR) values by 0.7?dB to 3.5?dB. Key words: image restoration; constrained optimization problem; sparse representation; alternating minimization; iterative soft-thresholding ? 0 引言? 图像成像模型可表示为：? ?y=Hx+n(1)? 或? y=HDz+n(2)? 其中：y∈R??M×1?表示已知的观测图像；x∈R??N×1?表示未知的原始图像；线性算子H∈R??M×N?；n∈R??M×1?N改为M表示加性噪声；D∈R??N×L?表示字典。特别地，若D是?Parseval?框架小波（即归一化的紧框架小波），它满足D??T?D=I［1］；z∈R??L×1?表示稀疏系数。式(1)和(2)分别代表了图像复原的两类主要算法：分析优化算法［2-4］和稀疏优化算法［5-7］。之所以将后一类算法称为稀疏优化算法，是因为该类算法首先估计向量z的最稀疏解，然后通过x=Dz重建原始图像。其思想来源于最近提出的压缩传感理论（?compressed sensing?）［8］，该理论认为：虽然空间域的图像大多是非稀疏的，但是选择合适的变换域，可得到任何信号的稀疏表示。除了图像复原领域，在图像降噪［9］、图像修复［10］、图像重建［11］等领域，图像的变换稀疏性也获得了成功应用。图像复原的稀疏优化问题可表示为：? ??min??