应用自适应权值矩阵的图像复原  

Image restoration using adaptive weight matrix

作　　者：朱齐丹[1] 孙磊[1] 蔡成涛[1] 

机构地区：[1]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《光学精密工程》2013年第6期1592-1597,共6页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.61175089;No.61203255);博士后研究人员落户黑龙江科研启动资金资助项目(No.LBH-Q11135);中央高校基本科研业务费专项资金资助项目(No.HEUCF1321001)

摘　　要：由于自然图像先验模型的概率密度函数的非高斯性(稀疏性的)会导致图像复原的最优化函数不再是凸的,传统意义上的最大后验估计已不能很好地求得最优估计解,因此本文提出利用自适应权值矩阵来解决这一问题。进行图像复原时,首先利用自然图像先验模型有效地抑制振铃效应,然后利用基于自适应权值矩阵的共轭梯度算法来解决由于稀疏先验模型导致的最优化函数非凸的问题。权值矩阵可根据上一次的迭代结果进行更新,并能够纠正图像在上一次迭代过程中局部区域导数估计的错误。实验结果显示,利用本文方法得到复原图像的峰值信噪比(PSNR)为36.131 6,优于其它算法。最后,用本文方法对全景图像进行复原,得到了很好的复原效果,证明了本文方法的实用性和有效性。Because natural image filter response probability model is often a non-Gaussian form （sparse）, it leads to the optimization problem of image restoration to be a non-convex and the optimal estimation solution can not be obtained by traditional maximum a posterior estimation. Therefore, this paper proposes a image restoration algorithm based on adaptive weight matrix to solve the prob- lem. With image restoration, a prior model for the natural image is used to restrain the ring effect ef- fectively and the conjugate gradient algorithm based on adaptive weight matrix is used to solve the problem of the non-convex optimization function due to the sparse prior. The weight matrix updates according to the last iteration result and is able to correct the error of the local image derivative esti- mation in the last iteration process. Experiments show that Peak Signal to Noise Ratio（ PSNR）gotten by proposed algorithm is 36. 131 6, better than that from other algorithms. Finally, the panoramic image is restored by proposed method and the good results are also obtained, which demonstrates that the algorithm proposed is practical and effective.

关 键 词：图像复原 反降晰 稀疏先验 自适应权值矩阵 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]