基于梯度L0稀疏正则化的图像盲去模糊算法

Image blind deblurring using L0 sparse regularization

图像的盲复原旨在从输入的模糊图像中估计出模糊核,并以此恢复出原清晰图像。本文分别采用近似L0和L1范数对待估清晰图像的梯度域和点扩散函数的稀疏性进行先验约束,后续的迭代过程中分别利用频域直接求解、分裂Bregman方法以及Shrinkage收缩阈值操作符进行优化。本文采用的表达形式既无需显性地在迭代过程中通过额外步骤提取出强梯度特征(例如冲击滤波器),又能保证在较少的迭代次数中达到收敛。最终通过实验结果验证了本文中采用的方法在图像恢复质量和迭代速度方面都有所改善。

Image blind deblurring aims to estimate the blur kernel store the latent sharp one, which is a typically ill-posed problem. In ble, many methods are proposed to formulate the cost function by paper, we introduce a framework applying LO and L1 norm sparsity from the input blurry order to make it more image and to re- robust and relia- representing the sparsity for the latent image and priors. In this blur kernel re- spectively. Effective numerical approaches including split Bregman and shrinkage thresholding methods are used in the optimization. The proposed scheme is proved to have better restoration quality and speed corn- paring with other methods used in the experiment.