Leveraging from the theory of
sparse signal representation and
sparse sampling (RecoSamp)
we develop registration
algorithms that can register
accurately regions of an image
using only a very small number
of features. Given a set of well
registered images we then
perform super-resolution using
appropriate sparsity priors.
There are many different image
models available that lead to
sparse representation. The one
we use assumes that the signal
is well approximated by a
piecewise polynomial image. This
model is ideal for depth images
but is also suitable for most
natural images. Our
approximation is calculated
using a quadtree decomposition
to adaptively partition the
image. This approach leads to
state of the art
super-resolution algorithms and
to state-of-the art image
denoising algorithms for very
noisy images.