1: \begin{abstract}

2: Patch-based low rank is an important prior assumption for image processing.

3:  Moreover, according to our calculation, the optimization of $l_0$ norm corresponds to the maximum likelihood estimation under random-valued impulse noise. In this article, we thus combine exact rank and $l_0$ norm  for removing the noise. It is solved formally using the alternating

4: direction method of multipliers (ADMM), with our previous patch-based weighted filter (PWMF) producing initial images. Since this model is not convex, we consider it as a Plug-and-Play ADMM, and do not discuss theoretical

5: convergence properties. Experiments show that this method has very good performance, especially for weak or medium contrast images.

6: %To do this, a key point is the detection of  impulse noise. %We first categorize the images, and use different process for different type of images.

7: \end{abstract}

8: