An Optimal Blind Temporal Motion

ranchocucamongabrrrΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 3 χρόνια και 7 μήνες)

53 εμφανίσεις

An Optimal Blind Temporal Motion

Blur Deconvolution Filter


The frames of a video sequence can be improved by a

spatial deconvolution of any motion
blur not exceeding two pixels

per frame. Yet, this requires an accurate blur estimation and

deconvolution, which is problematic formultiple local motions.We

introduce an optimal
temporal blur deconvolution filter restoring

blindly any nonuniform motion blur with an
amplitude below one

pixel per frame. The discrete filter has a very low complexity


20 operations per pixel. Experiments illustrate the method

on simulated data, real movies and
on sequences from the Middlebury



MAGE deconvolution is a classical inverse problem in

image processing. Deconvolution
appears in

a wide range

of application areas, such as photography, astronomy, medical

imaging, and remote sensing, just to name few. Images deteriorate

during acquisition as data
pass through the sensing,

transmission, and recording processes. In general, the observ

degradation is a result of two physical phenomena. The first is

of random nature and appears
in images as noise. The second is

deterministic and results in blurring, which is typically
by convolution with some blur kernel called the point sprea

function (PSF).
Degradation caused by convolution can thus

appear in any application where image
acquisition takes place.

The common sources of blurring are lens imperfections, air

turbulence, or camera
scene motion. Solving the deconvolution

problem in
a reliable way has
been of prime interest in the field

of image processing for several decades and has produced

enormous number of publications.

Let us first consider problems with just one degraded

i.e., single
channel deconvolution. The simples
t case is if the

blur kernel is known
(i.e., a classical deconvolution problem).

However, even here, estimating an unknown image
is ill

due to the ill
conditioned nature of the convolution operators.

This inverse
problem can only be solved by adoptin
g some

sort of regularization (in stochastic terms,
regularization corresponds

to priors). Another option is to use techniques such as

aperture , but this requires a modification of camera

hardware, which we do not consider


In the existing systed they have presented a algorithm for solving Multi channel(MC)
blind deconvolution. The existing system approach starts by defining an optimization
problem with image and blur regularization terms. To force sparse image gradients,

image regularizer is formulated using a standard isotropic TV.

The PSF regularizer consists of two terms:MCconstraint (matrix ) and sparsity

positivity. The MC constraint is improved by considering image Laplacian, which
brings better noise robustness at

little cost. Positivity helps the method to convergence
to a correct solution, when the used PSF size is much larger than the true one.

And solves the optimization problem in an iterative way by alternating between
minimization with respect to the image
step) and with respect to the PSFs (
Sparsity and positivity imply nonlinearity, but by using the variable splitting and
ALM (or split
Bregman method), we can solve each step efficiently, and moreover,
convergence of each step is guaranteed.


Synthetic data illustrate low


of the algorithm,


robustness to noise, and

stability in

the case of overestimated PSF sizes


We have proposed a new filter to process video sequences acquired by classic
cameras. By
applying the filter for a given maximal velocity, video sequences can be deblurred exactly
and automatically by a fast fixed temporal filter requiring only the knowledge of the maximal
observed velocity. This filter has been tested on synthetic

results and permits a decrease of
up to 20% in RMSE (i.e., dB).

The filter can be expected to work locally in all regions of the image where the motion is
lower than 1 pixel per frame. Nevertheless, a blind application seems to give good visual
in a wider range of velocities. This can be explained by the fast decay of the
frequency content of images. (Beyond 2 pixels per frame the deconvolution becomes ill
posed anyway.) An (exact) knowledge of the velocity

vector would also permit a direct
al (not temporal) deconvolution, but the temporal filter works without requiring this
accurate estimation.

These results mean that


it is possible to turn any camera

into a camera that ensures a sharp image for a broader
range of

velocities than a standar
d camera


it is possible to perform an

automatic blind movie deconvolution.

In the second case, the

proposed apparatus remains extremely simple as it creates the

sharp image as a linear combination of blurry images.

Therefore, the computational
cost is quasi
null. Furthermore,

assuming that the camera has provided a burst satisfying the

maximal velocity
condition, the proposed filter can be considered

an useful addition, not impeding other image



Image is denoi
sed by accumulation or spatial PSF deconvolution

High resoluted noiseless image


MATLAB 7.12(R2011a)