Natural or synthetic image can be decomposed into pattern images with regular or near regular objects. Effective separation makes possible to track object or recognize and recover hidden area from occlusion, or estimate the background from video. To separate high dimensional data with low rank matrix and sparse matrix, Robust Principal Component Analysis, RPCA is commonly used since it is stronger for gross error or outliers than PCA. There are many algorithms for convex optimization problem formulated by RPCA. Among them the Augmented Lagrange Multiplier Method are very fast and converge to exact optimal solution. This paper focuses on the regularization parameter dependent on input signal complexity, such as rank, instead of previous work has fixed value dependent on the size of row. The rank of input image helps to predict the weight of low rank matrix and sparse matrix. A number of experimental results prove that our regularization parameter is robust on various situations.