Illumination normalizing method and apparatus
View Patent ↗An illumination normalizing apparatus and a method are disclosed. The illumination normalizing apparatus measures a discontinuity of each pixel of an input image, the discontinuity including a spatial gradient and a local inhomogeneity, produces a weight of each pixel from the discontinuity by using a transfer function, produces an estimated illumination by repeating a convolution operation on each weight, and subtracts the estimated illumination from the input image.
1. An illumination normalizing apparatus, comprising:
a discontinuity measuring means, measuring a discontinuity of each pixel of an input image, the discontinuity comprising a spatial gradient and a local inhomogeneity;
a weight calculating means, producing a weight of each pixel from the discontinuity by using a transfer function;
an illumination estimating means, producing an estimated illumination by repeating a convolution operation on each weight; and
an illumination normalizing means, subtracting the estimated illumination from the input image,
in which the discontinuity measuring means comprises:
a gradient measuring means, producing the spatial gradient by performing a partial differential on the pixel; and
an inhomogeneity measuring means, producing differences between luminance of the pixel and each luminance of n adjacent pixels and the mean of differences,
in which the discontinuity measuring means controls the gradient measuring means and the inhomogeneity measuring means to be independently operated in parallel,
in which the inhomogeneity measuring means produces the local inhomogeneity by the following formula:
τ
~
(
x
,
y
)
=
sin
(
π
2
τ
s
(
x
,
y
)
)
wherein
τ
s
(
x
,
y
)
is
τ
(
x
,
y
)
-
τ
min
τ
max
-
τ
min
,
τ
(
x
,
y
)
is
∑
∑
(
m
,
n
)
∈
Ω
I
(
x
,
y
)
-
I
(
m
,
n
)
Ω
,
τ max and τ min are the maximum value and the minimum value among τ(x, y), respectively, Ω indicates k adjacent pixels that are adjacent to the pixel at (x, y), (m, n) is a coordinate of an adjacent pixel included in Ω, and m and n are integers;
in which the weight is produced by the following formula:
w
(
x
,
y
)
=
1
1
+
τ
~
(
x
,
y
)
h
×
1
1
+
∇
I
(
x
,
y
)
S
wherein h and S are real numbers.
2. The illumination normalizing apparatus of claim 1 , in which the gradient measuring means produces the spatial gradient by the following formula:
|∇ I ( x,y )|=√{square root over ( G x 2 +G y 2 )}
wherein I(x, y) is a value of a pixel that locates on a coordinate (x, y), and
G x =I ( x+ 1 ,y )− I ( x −1, y )
G y =I ( x,y +1) −I ( x,y −1)
wherein x and y are integers equal or greater than 0.
3. The illumination normalizing apparatus of claim 1 , in which the convolution operation is performed by the following formula:
N
(
x
,
y
)
(
t
)
=
∑
i
=
-
1
1
∑
j
=
-
1
1
w
(
t
)
(
x
+
i
,
y
+
j
)
L
(
t
+
1
)
(
x
,
y
)
=
max
{
1
N
(
x
,
y
)
(
t
)
∑
i
=
-
1
1
∑
j
=
-
1
1
L
(
t
)
(
x
+
i
,
y
+
j
)
w
(
t
)
(
x
+
i
,
y
+
j
)
,
L
(
t
)
(
x
,
y
)
}
wherein L (t) is an estimated illumination of each pixel when the convolution operation is performed t times, w (t) (x, y) is a weight of each pixel when the convolution operation is performed t times, and i and j are integers.
4. The illumination normalizing apparatus of claim 1 , in which the illumination normalizing means performs a subtraction between the logarithm of the input image and the logarithm of the estimated illumination, and normalizes the result of subtraction.
5. An illumination normalizing method, comprising:
producing at least one discontinuity for each pixel of an input image;
producing a weight of each pixel from the discontinuity by using a transfer function;
producing an estimated illumination by repeating a convolution operation on each weight; and
subtracting the estimated illumination from the input image,
in which the discontinuity comprises a spatial gradient and a local inhomogeneity,
in which the spatial gradient is produced by the following formula:
|∇ I ( x,y )|=√{square root over ( G x 2 +G y 2 )}
wherein I(x, y) is a value of a pixel that locates on a coordinate (x, y), and
G x =I ( x +1 ,y )− I ( x −1 ,y )
G y =I ( x,y +1) −I ( x,y −1)
wherein x and y are integers equal or greater than 0,
in which the local inhomogeneity is produced by the following formula:
τ
~
(
x
,
y
)
=
sin
(
π
2
τ
s
(
x
,
y
)
)
wherein
τ
s
(
x
,
y
)
is
τ
(
x
,
y
)
-
τ
min
τ
max
-
τ
min
,
τ
(
x
,
y
)
is
∑
∑
(
m
,
n
)
∈
Ω
I
(
x
,
y
)
-
I
(
m
,
n
)
Ω
,
τ max and τ min are the maximum value and the minimum value among τ(x, y), respectively, Ω indicates k adjacent pixels that are adjacent to the pixel at (x, y), (m, n) is a coordinate of an adjacent pixel included in Ω, and m and n are integers,
in which the weight is produced by the following formula:
w
(
x
,
y
)
=
1
1
+
τ
~
(
x
,
y
)
h
×
1
1
+
∇
I
(
x
,
y
)
S
wherein h and S are real numbers.
6. The illumination normalizing method of claim 5 , in which the convolution operation is performed by the following formula:
N
(
x
,
y
)
(
t
)
=
∑
i
=
-
1
1
∑
j
=
-
1
1
w
(
t
)
(
x
+
i
,
y
+
j
)
L
(
t
+
1
)
(
x
,
y
)
=
max
{
1
N
(
x
,
y
)
(
t
)
∑
i
=
-
1
1
∑
j
=
-
1
1
L
(
t
)
(
x
+
i
,
y
+
j
)
w
(
t
)
(
x
+
i
,
y
+
j
)
,
L
(
t
)
(
x
,
y
)
}
wherein L (t) is an estimated illumination of each pixel when the convolution operation is performed t times, w (t) (x, y) is a weight of each pixel when the convolution operation is performed t times, and i and j are integers.
7. The illumination normalizing method of claim 5 , in which the subtracting the estimated illumination from the input image is a subtraction between the logarithm of the input image and the logarithm of the estimated illumination, and normalizes the result of subtraction.
8. A non-transitory computer-readable storage medium including a program containing computer-executable instructions for illumination normalization performing the method comprising:
producing a vector of spatial gradient for each pixel of an input image;
measuring a local inhomogeneity that indicates the degree of inhomogeneity between the pixel and adjacent pixels;
producing a weight for retaining a feature of the input image and for removing a shadow having discontinuity similar to the feature by use of the spatial gradient and the local inhomogeneity;
producing an estimated illumination by repeating a convolution operation by use of the weight; and
performing a subtraction between the logarithm of the input image and the logarithm of the estimated illumination,
in which the discontinuity comprises a spatial gradient and a local inhomogeneity,
in which the spatial gradient is produced by the following formula:
|∇ I ( x,y )|=√{square root over ( G x 2 +G y 2 )}
wherein I(x, y) is a value of a pixel that locates on a coordinate (x, y), and
G x =I ( x +1 ,y )− I ( x −1 ,y )
G y =I ( x,y +1) −I ( x,y −1)
wherein x and y are integers equal or greater than 0,
in which the local inhomogeneity is produced by the following formula:
τ
~
(
x
,
y
)
=
sin
(
π
2
τ
s
(
x
,
y
)
)
wherein
τ
s
(
x
,
y
)
is
τ
(
x
,
y
)
-
τ
min
τ
max
-
τ
min
,
τ
(
x
,
y
)
is
∑
∑
(
m
,
n
)
∈
Ω
I
(
x
,
y
)
-
I
(
m
,
n
)
Ω
,
τ max and τ min are the maximum value and the minimum value among τ(x, y), respectively, Ω indicates k adjacent pixels that are adjacent to the pixel at (x, y), (m, n) is a coordinate of an adjacent pixel included in Ω, and m and n are integers,
in which the weight is produced by the following formula:
w
(
x
,
y
)
=
1
1
+
τ
~
(
x
,
y
)
h
×
1
1
+
∇
I
(
x
,
y
)
S
wherein h and S are real numbers.