Fast image deconvolution using Hyper-Laplacian Prior
Dilip Krishnan Rob Fergus
New york University
Presented by Zhengming Xing
Outline
• Introduction
• Algorithm
• Experiment result
introduction
• Hyper-Laplacian Prior p ( x )
exp(
k | x |
) ( typically with 0 .
5
0 .
8 )
• speed
algorithm
For non-blind deconvolution problem
Given y (the blurred image), and k( blur kernel), x(original image). Assume Gaussian noise.
p ( x | y , k )
p ( y | x , k ) p ( x )
Hyper-Laplacain prior
Minimize
log p ( x | y , k )
Optimize problem recall
Half quadratic penalty method, introduce auxiliary variable.And consider the one special case.
j
2 .
f
1
[ 1
1 ] , f
2
[ 1
1 ]
T
F i j x
( x
f j
) i
Recall:
Solve sub-problem
• Fixed w
Recall:
Solve sub-problem
Fixed X
Lookup table: pre-compute solution for different
Analytic solution: for particular value of
,
Recall:
Take derivative
1 / 2
2 / 3
Compare the different root and find the global minimum
Summary of the algorithm
Summary of the algorithm
Experiment description
• Grey scale real world image, blurred by camera shaked kernels and add Gaussian noise. The kernels are minor perturbed.
• Measured with the SNR
10 log
10
(
|| x
ˆ
||
x
ˆ
( x
ˆ
) x ||
2
||
2
)
result
result
result