The Weizmann Institute of Science
Faculty of Mathematics and Computer Science
Vision and Robotics Seminar
Matan Protter
Technion
will speak on
Super-Resolution without explicit motion estimation
Abstract:
Super-resolution is the task of improving the optical resolution of a given
low-quality image sequence. The motion in the sequence introduces different
aliasing artifacts in each frame, which are then used by super-resolution
algorithms to recover sub-pixel details. Since the motion is important for
recovering sub-pixel details, it has been long assumed that motion estimation
is a necessary step for any super-resolution algorithm. Indeed, when the motion
is known, amazing results can be achieved. However, due to the complexity of
motion estimation and the high accuracy required of it, successful
super-resolution has been limited to processing sequences containing only
global motion (such as translations) or very simple motion patterns.
In this work, we challenge the assumption that explicit motion estimation is
indeed required. Instead, we propose a method that avoids explicit motion
estimation, relying on what could be interpreted as fuzzy motion estimation, in
the sense that for each pixel several trajectories are considered and merged
probabilistically.
We put the fuzzy motion concept into use in two ways. This final algorithm we
develop is surprisingly simple, and is very reminiscent of (but critically
different from) the Non-Local-Means image sequence denoising algorithm. We
display several results on sequences with general and complicated motion
patterns, proving the ability of the proposed method to process such movies and
provide true super-resolution outcome.
Joint work with Michael Elad (Technion), Peyman Milanfar (UCSC) and Hiro Takeda
(UCSC)
The lecture will take place in the
Lecture Hall, Room 1, Ziskind Building
on Thursday, April 3, 2008
12:00 - 13:00