MRI Super-Resolution with Two Regularization Parameters

The artifacts and noise in the recovered images by the regularization super-resolution (SR) algorithms based on sparse coding are obvious. A proposed SR algorithm for MRI images via two regularization parameters can improve the SR performance in this paper. With the hypothesis that the sparse coefficients in the HR space and LR space are different in the dictionary training phase while the sparse coefficients in the two spaces are the same in the image reconstruction phase. Based on this, the algorithm in this paper introduces online dictionary learning algorithm to generate the accurate dictionary pair separately with the training regularization parameter lambda(t). In the image reconstruction phase, the reconstruction regularization parameter lambda(r) is tuned to solve the best reconstruction sparse coefficient to recover the predicted HR image. In the experiments, the average PSNR and SSIM of the reconstructed images by the proposed algorithm is 1.3dB and 0.023 higher than the typical Couple Dictionary Learning SR algorithm. The SR performance is raised considerably to eliminate the noise and artifacts effectively.