Upload
clifton-flowers
View
218
Download
0
Tags:
Embed Size (px)
Citation preview
1
Removing Camera Shake from a Single Photograph
Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T. Roweis and William T. Freeman
ACM SIGGRAPH 2006, Boston, USA
4
Image formation process
=
Blurry image Sharp image Blur kernel
Input to algorithm Desired output
Convolutionoperator
+
Gaussian noise
5
Image formation process
A linear imaging model is assumed in this paper, that is:
where
: Desired image
: Observation
: Blur kernel
: Gassian noise
Posterior probability :
( , | )} ( | , ) ( , ) ( | , ) ( ) ( )
B K L N
L
B
K
N
p L K B p B L K p L K p B L K p L p K
Prior for blur kernel
Prior for image
Observation model
6
Prior model for nature image (1)
Characteristic distribution of sharp image with heavy tails
L
Histogram of image gradient ( )L i
7
Prior model for nature image (2)
Use parametric model for sharp image statistics
L
1
Mixture of Gaussian model for the gradient at pixel i :
( ( )) ( ( ) | 0, )C
c cc
p L i N L i v
1
Pdf for image gradients :
( ) ( ( ) | 0, )C
c cci
p L N L i v
8
Prior model for blur kernel (1)
The characteristics of blur kernel are positive and sparse
=
Blurry image Sharp imageBlur
kernel
9
Prior model for blur kernel (2)
Assume the probability distribution of the element of blur kernel is the mixture of exponential distributions
Exponential distribution
1
1
Mixture of Exponential model
for kernel element:
( ) ( | )
Pdf for blur kernel:
( ) ( | )
D
j d j dd
D
d j ddj
p K E K
p K E K
10
Model transformation
The imaging model need to be transformed before we using the image gradient prior, that is:
2
11
Original formulation:
( , | ) ( | , ) ( ) ( )
Modified formulation:
'
( | , )
{ ( ( ) | (
( )
{
( ( )
{ ( ( ) | 0, )}
, | )
), ( | )= ) *}C
c ccii
D
d j dd
B K L N
p L K B p B L K p L p K
B K L N
p p p K
E K
B L K L
N L i
p L K
N B i K v
B
L i
} j
11
Variational Baye (1)
Illustration for Bayesian mean squared error estimator (Minimum mean squared error estimator, MMSE)
Ө1
Ө2
Ө3Process
Parameter space Ө
d
Observed data
1 1 2 2 3 3
MMSE for (Weighted aveage):
* ( | ) * ( | ) * ( | )p d p d p d
12
Variational Baye (2)
Apply MMSE estimator for blur kernel estimation
K
K
MMSE for Blur kernel:
' * ( | )
{ ( , | ) }
dd dL
K K p K B
K p L K B
It may be difficult to find the integration result of the posterior probability
13
Variational Baye (3)
Factorize the previous posterior probability for blur kernel inference operation
( , | )p L K B
K K
K K ( )
The factorization of posterior probability:
( , | ) ( ) ( )
MMSE for Blur kernel now becomes:
' * ( | ) { ( , | ) }
{ ( ) ( ) } ( )
d d dd d d
L
L q K
p L K B q K q L
K K p K B K p L K B
K q K q L Kq K K
14
Variational Baye (4)
The factorization of the posterior probability could be modeled as an optimization problem
, , ,
, ,
Assume
( ) ~ ( ; , )
( ) ~ ( ; , )
where
, , , are unknown parameters
Distribution approximation:
', ', ', ' arg min Distance( ( ) ( ), ( , | ))
arg minK K L L
K K
L L
rectified K K
L L K K
K K L L m v m v
m v m
q L N L m v
q K N K m v
m v m v
m v m v q L q K p L K B
,
_ ( ( ) ( ), ( , | ))L LvKL divergence q L q K p L K B