View the article online for updates and enhancements. Abstract. In this paper we propose an −subgradient method for solving a constrained minimization problem arising in super-resolution imaging applications. The method, compared to the state-of-the-art methods for single image super-resolution on some test problems, proves to be very efficient, both for the reconstruction quality and the computational time.
IntroductionImage super-resolution reconstruction is the process of obtaining High Resolution (HR) images from observed Low Resolution (LR) images. The problem of super-resolution is of great importance in many applications, such as medicine or object recognition (face, bar codes, ...), where the electronic imaging devices are equipped with low resolution cameras, while HR images are finally desired. The super-resolution process can be performed from a single image (SISR) or from multiple images of the same scene (MISR), such as in the case of videos or medical imaging devices for example. In this paper, we will consider only the case of SISR. The SIRS problem is ill-posed, since identical LR images can be generated from different HR images. Furthermore, in addition to being connected by a down-sampling operator, the HR and LR images are related by a Point Spread Function (as a simple convolution, for example), that is an ill-posed operator. The different approaches, present in literature for image super-resolution, can generally be grouped into four categories: interpolation-based algorithms, example-based algorithms, sparse-representation-based algorithms and reconstruction-based algorithms (see [5] and the references therein). The reconstruction-based algorithms are aimed at solving a minimization problem, where the objective function is the sum of a fidelity term related to the data noise and a regularization term, representing the prior on the solution and, at the same time, enabling to face the ill-posedness of the problem. This formulation has the advantage of simultaneously exploiting the degradation model and the information contained in the prior, thus reducing noise and artifacts in the HR reconstructed image [5,12].We consider a reconstruction-based algorithm to address a minimization problem where the fidelity term is the 2 -norm, the regularization term is the discrete Total Variation (TV) function and the solution is constrained to be nonnegative. This model has been widely treated for the deblurring applications and many algorithms have been proposed for its solution, such as, for example, [11,13,14]. The Bregman method proposed in [14] has been later used in [12] for image super-resolution. In this paper we propose a scaling -subgradient method presented in [2] for deblurring of images corrupted by Poisson noise. The method has a fast computational