SR image enhancement approaches have received a lot of
interest in the past two decades, and a variety of SR algorithms
have been proposed in the literature. The idea of SR was
first introduced in 1984 by Tsai and Huang for LANDSAT4
images [4]. Kim et al. generalized this work to noisy and
blurred images, using least square minimization [5]. In both
of these studies, SR is performed in the frequency domain.
Although frequency-based SR methods have the advantage of
simplicity and low computational complexity, they require that
the geometric disparity between LR images is translational.
In order to overcome this problem, various spatial-based SR
approaches have been proposed as they provide more flexibility
in modeling the image degradation process. A SR spatial
domain algorithm was first presented by Ur and Gross in 1992
[6]. Utilizing the generalized multichannel sampling theorem
they proposed a non-uniform interpolation method for multiple
spatially shifted LR images. The interpolation is followed by a
deblurring process. A different method called Iterative Backward
Projection (IBP), which was adopted as a basic algorithm
in computer-aided tomography, was first suggested by Irani and
Peleg [7]. If the image degradation elements including blurring,
warping, sampling, and additive noise are known exactly,
SR becomes an inverse problem similar to image restoration,
which has been studied for a much longer time than SR. Many
algorithms coming from the image restoration domain can then
be used for SR, and in some way SR can be thought of as
second-generation image restoration. To reduce noise and solve
singular cases, several SR algorithms incorporate prior knowledge
into the computation by constraining the solution. For
example, Li et al. proposed the maximum a posteriori based
method (MAP) using a discontinuity preserving universal
Hidden Markov Tree (HMT) model for SR reconstruction [8].
For an overview of state-of-the-art SR methods, we refer to [2].
Not all LR images, however, can be employed for SR image
enhancement. First, imagesmust be undersampled, that is to say,