plug-and-play;出处:CVPR2019
Method
plug-and-play image restoration:
[15]. Aram Danielyan, Vladimir Katkovnik, and Karen Egiazarian. Image deblurring by augmented lagrangian with BM3D frame prior. In Workshop on Information Theoretic Methods in Science and Engineering, pages 16–18, 2010.
[57]. Singanallur V Venkatakrishnan, Charles A Bouman, and Brendt Wohlberg. Plug-and-play priors for model based reconstruction. In IEEE Global Conference on Signal and Information Processing, pages 945–948, 2013.
[69]. Daniel Zoran and Yair Weiss. From learning models of natural image patches to whole image restoration. In IEEE International Conference on Computer Vision, pages 479–486, 2011.
做法:(1). 通过变量分离展开能量方程;(2). 把和先验相关的子问题换成任意现有的高斯去噪器。
plug-and-play相关工作的进展:
- 不同的变量分离算法:half-quadratic splitting (HQS), alternating direction method of multipliers (ADMM), FISTA, primal-dual …
- 不同任务:Poisson denoising, demosaicking,deblurring,super-resolution,inpainting
- 不同的denoiser prior:BM3D, DNN-based denoiser (通常用Gaussian denoiser)
- 从不动点与纳什均衡两方面对收敛性的理论分析
为了方便估计kernel,将degradation model写成:
这里的$\downarrow_s$是bicubic下采样。所以SR问题变为了先做SISR再做deblurring。
根据最大后验概率(MAP)写出能量方程(energy function):
引入辅助变量$\mathbf{z}$, 求解:
下面用HQS求解:HQS
得到迭代解:
得到闭式解:
对比式(10)和能量方程,$\mathbf{z}_{k+1}$由$\mathbf{x}$ 先bicubic下采样s倍,再由噪声level为$\sqrt{1/\mu}$的AWGN得到。