Robust super-resolution reconstruction with adaptive regularization

In the last two decades, super-resolution reconstruction is an active topic in image and video processing which is theoretically important as well as practically urgent in many fields. There are a variety of methods for super-resolution reconstruction such as Bayesian maximum a-posteriori (MAP), weighted least square (WLS) and projection onto convex sets (POCS) etc. Unfortunately, these methods are usually very sensitive to their assumed models of data and noise, which limits their utility. In this paper, a robust super-resolution reconstruction method for image sequences is proposed. Firstly, some different robust maximum likelihood estimators are introduced to consist the data fitting term. On the other hand, to overcome the ill-posed problem of maximum likelihood estimation, a robust regularization term is added, and results in reconstructed image with sharp edges. Furthermore, we propose the use of regularization functional instead of a constant regularization parameter. The regularization functional is defined in terms of the reconstructed image at each iteration step, therefore allowing for the simultaneous determination of its value and the reconstruction of the super-resolution image. The iteration scheme, convexity and control parameter are thoroughly studied. Experimental results demonstrate the power of the proposed method.