Compressed sensing MRI (CS-MRI) has demonstrated its potential in image reconstruction from undersampling k-space data to accelerate scanning time. Exact CS-MRI reconstruction is based on the image prior information of sparsity, and therefore, it relies on two important aspects of sparse domain and sparse coding. In this work, a patch based uniform model according to orthogonal dictionary learning and lp norm minimization (ODNM) is developed. First, image patches are classified into multi-classes by a use of a modified clustering method and the corresponding orthogonal dictionary is learned for each class. In addition, the non-convex lp norm regularization is employed to promote the sparsity of patch coefficients. In order to solve the proposed reconstruction model, the alternating direction method (ADM) steps are developed in which orthogonal dictionary learning, non-convex sparse coding, and image reconstruction are simultaneously conducted. In the special case of p=0.5, a fast shrinkage operator is proposed to reduce the computational complexity. Extensive experiments on real complex valued MR images under various sampling patterns demonstrate that the proposed method achieves a better performance than several state-of-the-art algorithms.