In video surveillance, microscopy, image restoration and many other applications, High-Resolution (HR) images are required
when only Low-Resolution (LR) images are available. This situation arises when details to be analyzed are beyond the
resolution of the imager.
Multi-Frame Super-Resolution (MFSR) is a way to overcome
this problem. It consists in reconstructing an HR
image from one or several overlapping LR images of
the same scene acquired with the same imager. This
technique relies on the fact that several images acquired
from the same scene contain different information
when there are slight motions between the
images.
To achieve MFSR, LR images have to be sub-pixel
registered and merged. Then, a deconvolution process
enables the reversal of the blur induced by the
so-called Point Spread Function (PSF) of the imager.
Until now, this last procedure relied heavily on the
precise modeling of the imager’s PSF. However, identifying
this PSF is nearly impossible.
The proposed method does not need such a model. It
uses the new mathematical tool we developed, called
maxitive kernels, to achieve an imprecise modeling of
the PSF based on very basic and generic information
about the imager.
One particularity of this method is its ability to produce
an interval-valued reconstructed HR image. We
have proposed a-posteriori regularization for selecting
the most representative image in this convex set
of HR images. This method possesses a robustness
advantage over its competitors with respect to modeling,
registration, overfitting and artifacts.
[1] F. Graba, F. Comby, O. Strauss: Non-Additive Imprecise Image Super-Resolution in a Semi-Blind Context. IEEE Trans. Image Processing 26(3): 1379-1392 (2017)