Denoise method on image pyramid
View Patent ↗The present invention is to provide a denoise method on Gaussian/Laplacian image pyramid, which integrates Pyramid analysis/synthesis algorithm, MMSE (minimum mean square error) filter and NL (non local) filter on the image pyramid to reconstruct and output a denoised image of an original input image through a plurality of iterative procedures, and utilizes an auto-adaptive noise estimation algorithm to find parameter of noise level used by the NL filter, so as to be easily embedded in mobile or handheld devices for obtaining better noise removing and anti-shaking results and remove noise much faster than the conventional denoise method, but only with less quality loss.
1. A denoise method on image pyramid, which is implemented to mobile or handheld devices, comprising:
a pyramid analysis/synthesis algorithm, comprising pyramid analysis and pyramid synthesis in a plurality of iterative procedures respectively wherein, in each iterative procedure, the pyramid analysis is used to create a Level j+1 approximation image by down-sampling a Level j image, create a first Level j prediction image by up-sampling the Level j+1 approximation image through using a first interpolation, and compute the difference between the first Level j prediction image and the Level j image for obtaining a level j frequency image, and the pyramid synthesis is used to create a second Level j prediction image by up-sampling a Level j+1 denoised image and create a Level j denoised image by adding the second Level j prediction image to an output of a noise filter which is used to filter the level j frequency image;
non local (NL) filter and minimum mean square error (MMSE) filter, included in the noise filter in each iterative procedure for removing gauss noise existing in the level j frequency image; and
an auto-adaptive noise estimation algorithm in each iterative procedure for finding parameter of noise level used by the NL filter.
2. The method of claim 1 wherein NL filter is denoted as {w m,n }:
w
i
,
j
=
1
M
exp
(
A
i
,
j
-
A
p
,
q
2
δ
2
)
where
M
=
∑
i
,
j
∈
Ω
exp
(
A
i
,
j
-
A
p
,
q
2
δ
2
)
so that
∑
i
,
j
w
i
,
j
=
1
;
δ is a parameter to denote the noise level of point (p,q); A i,j , A p,q are compare windows and small neighbor of (i,j) and (p,q) to compare the similarity of I i,j and I p,q , {I i,j } is the original input image, {I i,j *} is the filtered image, thus
I
p
,
q
*
=
∑
(
i
,
j
)
∈
Ω
w
i
,
j
I
i
,
j
here Ω is a neighbor of (p,q) named as search window.
3. The method of claim 2 wherein the compare window A i,j is a rectangle window with radius 3 and center (i, j) that:
A
p
,
q
-
A
i
,
j
2
=
∑
m
,
n
∈
[
-
3
,
3
]
V
p
+
m
,
q
+
n
-
V
i
+
m
,
j
+
n
2
here {V i,j } is a reference image to compute weight.
4. The method of claim 3 wherein {I i,j } is the frequency image to each level, and {V i,j } is the corresponding approximation image to use color message.
5. The method of claim 2 wherein the Search window Ω is selected as a rectangle window with radius from 3 to 12.
6. The method of claim 1 wherein the MMSE filter is denoted as:
f
*
(
p
)
=
{
f
_
(
p
)
+
Var
(
p
)
-
N
(
p
)
Var
(
p
)
(
f
(
p
)
-
f
_
(
p
)
)
Var
(
p
)
>
N
f
_
(
p
)
Var
(
p
)
≤
N
where f(p) is current pixel value on image to be filtered; f*(p) is the filtered pixel value; f (p) is the average value of a neighbor of pixel p; Var(p) is the variance of the neighbor of pixel p; and N(p) is the expected noise value of pixel p.
7. The method of claim 1 wherein the parameter of noise level is a noise level function (NLF) estimated by the auto-adaptive noise estimation algorithm comprising the steps of:
scanning the image and creating a two-dimensional Luminance-Variance distribution (L-V) table, wherein the L-V table is denoted as α[I][V] with Iε[0,255], Vε[0, V Max ], where I is the luminance level and V is the variance;
creating an initial NLF from the L-V Table; and
correcting the initial NLF to obtain a smooth line representing a resulted NLF.
8. The method of claim 7 , when creating the L-V table, further comprising the steps of:
initializing all cell of the L-V table to 0;
procuring each pixel on the approximation image and frequency image respectively;
computing the average I of the neighbor of current pixel procured from the approximation image;
computing the variance V of the neighbor of current pixel procured from the frequency image; and
increasing by one pixel at the corresponding cell α[I][V], and determining whether or not the computed variance V is more than V Max ; when the determination is positive, simply ignoring the computed variance V and continuing the subsequent steps; otherwise, looping back to procure next pixel.
9. The method of claim 8 , when creating the initial NLF, further comprising the steps of:
procuring each average I from 0 to 255;
finding background noise variance V or the maximal cell value α[I][V] from array {α[I][V]} at each luminance level I; and
increasing by one luminance level I, and then determining whether or not the luminance level I is more than 255; if yes, obtaining an initial NLF simply by selecting the variance value with max block number on each Luminance and continuing the subsequent steps; otherwise, looping back to procure next luminance level I.
10. The method of claim 9 , when correcting the initial NLF to obtain the resulted NLF, further comprising the steps of:
finding the main background noise variance V 0 at luminance I 0 ; and
finding the point in the L-V table with the maximum pair on the initial NLF by using {I 0 ,V 0 };
canceling other points on the initial NLF which differ too much to the maximum pair; and
interpolating the canceled points back to the L-V table by using the remained points to obtain the smooth line representing the resulted NLF.