Summary of the invention
Technical problem to be solved by this invention provides a kind of stereo image quality method for objectively evaluating that can effectively improve the correlation of objective evaluation result and subjective perception.
The present invention solves the problems of the technologies described above the technical scheme that is adopted: a kind of stereo image quality method for objectively evaluating based on visually-perceptible is characterized in that may further comprise the steps:
1. make S
OrgUndistorted stereo-picture for original makes S
DisFor the stereo-picture of distortion to be evaluated, with S
OrgLeft visual point image be designated as { L
Org(x, y) }, with S
OrgRight visual point image be designated as { R
Org(x, y) }, with S
DisLeft visual point image be designated as { L
Dis(x, y) }, with S
DisRight visual point image be designated as { R
Dis(x, y) }, wherein, (x, the y) coordinate position of pixel in left visual point image of expression and the right visual point image, 1≤x≤W, 1≤y≤H, W represent the width of left visual point image and right visual point image, H representes the height of left visual point image and right visual point image, L
Org(x, y) expression S
OrgLeft visual point image { L
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), R
Org(x, y) expression S
OrgRight visual point image { R
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), L
Dis(x, y) expression S
DisLeft visual point image { L
Dis(x, y) } in coordinate position be (x, the pixel value of pixel y), R
Dis(x, y) expression S
DisRight visual point image { R
Dis(x, y) } in coordinate position be (x, the pixel value of pixel y);
2. utilize the visual masking effect of human vision, extract undistorted left visual point image { L respectively background illumination and texture
Org(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image, with undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image be designated as { J
L(x, y) }, with undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image be designated as { J
R(x, y) }, wherein, J
L(x, y) expression { J
L(x, y) } in coordinate position be (x, the pixel value of pixel y), J
R(x, y) expression { J
R(x, y) } in coordinate position be (x, the pixel value of pixel y);
3. obtain undistorted left visual point image { L respectively through regional detection algorithm
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) }, undistorted right visual point image { R
Org(x, y) } and the right visual point image { R of distortion
Dis(x, y) } in the block type of each 8 * 8 sub-piece, be designated as p, wherein, { 1,2,3,4}, p=1 represent strong edge block to p ∈, and p=2 representes weak edge block, and p=3 representes smooth block, and p=4 representes texture block;
4. according to undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) }, through the left visual point image { L of calculated distortion
Dis(x, y) } in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects picture quality and the space structure intensity that is used to reflect picture quality and the right visual point image { R of distortion
Dis(x, y) } in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects picture quality and the space structure intensity that is used to reflect picture quality, obtain the left visual point image { L of distortion respectively
Dis(x, y) } be used to reflect the characteristic vector of picture quality and the right visual point image { R of distortion
Dis(x, y) } the characteristic vector that is used to reflect picture quality, again to the left visual point image { L of distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } be used to reflect that the characteristic vector of picture quality carries out linear weighted function, obtains S
DisThe characteristic vector that is used to reflect picture quality, be designated as F
q
5. according to undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) }, through the left visual point image { L of calculated distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } absolute difference image in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects depth perception and the space structure intensity that is used to reflect depth perception, obtain S
DisThe characteristic vector that is used to reflect depth perception, be designated as F
s
6. with S
DisThe characteristic vector F that is used to reflect picture quality
qWith the characteristic vector F that is used to reflect depth perception
sForm new characteristic vector, as S
DisCharacteristic vector, be designated as X, X=[F
q, F
s], " [] " is the vector representation symbol, [F
q, F
s] represent characteristic vector F
qWith characteristic vector F
sCouple together and form a new characteristic vector;
7. adopt n undistorted stereo-picture, set up its distortion stereo-picture set under the different distortion levels of different type of distortion, this distortion stereo-picture set comprises the stereo-picture of several distortions; Utilize the subjective quality evaluation method to obtain the average subjective scoring difference of the stereo-picture of every width of cloth distortion in the set of distortion stereo-picture respectively, be designated as DMOS, DMOS=100-MOS; Wherein, MOS representes the subjective scoring average, DMOS ∈ [0; 100], n >=1;
8. adopt and calculating S
DisThe identical method of characteristic vector X, the characteristic vector of the stereo-picture of every width of cloth distortion in the set of calculated distortion stereo-picture respectively, the characteristic vector for the stereo-picture of i width of cloth distortion in the set of distortion stereo-picture is designated as X with it
i, wherein, 1≤i≤n ', the width of cloth number of the stereo-picture of the distortion that comprises in the set of n ' expression distortion stereo-picture;
9. adopt support vector regression that the characteristic vector of the stereo-picture of all distortions of identical type of distortion in the set of distortion stereo-picture is trained; And the support vector regression training pattern of utilizing training to obtain is tested the stereo-picture of every width of cloth distortion of same type of distortion; Calculate the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion of identical type of distortion in the set of distortion stereo-picture; Evaluating objective quality predicted value for the stereo-picture of i width of cloth distortion in the set of distortion stereo-picture is designated as Q with it
i, Q
i=f (X
i), f () is the function representation form, Q
i=f (X
i) expression Q
iBe X
iFunction, wherein, 1≤i≤n ', the width of cloth number of the stereo-picture of the distortion that comprises in the n ' expression distortion stereo-picture set.
Described step detailed process 2. is:
2.-1, calculate undistorted left visual point image { L
Org(x, y) } the visual threshold value set of visual masking effect of background illumination, be designated as { T
l(x, y) },
Wherein, T
l(x, y) the undistorted left visual point image { L of expression
Org(x, y) } in coordinate position be (x, the visual threshold value of the visual masking effect of the background illumination of pixel y),
Represent undistorted left visual point image { L
Org(x, y) } in be that (x, pixel y) they are the average brightness of all pixels in 5 * 5 windows at center with coordinate position;
2.-2, calculate undistorted left visual point image { L
Org(x, y) } the visual threshold value set of visual masking effect of texture, be designated as { T
t(x, y) }, T
t(x, y)=η * G (x, y) * W
e(x, y), wherein, T
t(x, y) the undistorted left visual point image { L of expression
Org(x, y) } in coordinate position be that (η is the controlling elements greater than 0 for x, the visual threshold value of the visual masking effect of the texture of pixel y), and (x y) representes undistorted left visual point image { L G
Org(x, y) } in coordinate position be that (x, pixel y) carry out the maximum weighted mean value that directed high-pass filtering obtains, W
e(x, y) expression is to undistorted left visual point image { L
Org(x, y) } edge image in coordinate position be that (x, pixel y) carry out the edge weighted value that Gauss's LPF obtains;
2.-3, to undistorted left visual point image { L
Org(x, y) } the visual threshold value set { T of visual masking effect of background illumination
l(x, y) } and the visual threshold value set { T of the visual masking effect of texture
t(x, y) } merge, obtain undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image, be designated as { J
L(x, y) }, J
L(x, y)=T
l(x, y)+T
t(x, y)-C
L, t* min{T
l(x, y), T
t(x, y) }, wherein, C
L, tThe parameter of the visual masking effect eclipse effect of expression control background illumination and texture, 0<C
L, t<1, min{} is for getting minimum value function;
2.-4,2.-1 employing and step to 2.-3 identical operations, obtain undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image, be designated as { J
R(x, y) }.
The detailed process of the regional detection algorithm of described step in 3. is:
3.-1, respectively with undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } be divided into
8 * 8 sub-pieces of individual non-overlapping copies define undistorted left visual point image { L
Org(x, y) } in l 8 * 8 sub-pieces be the current first sub-piece, be designated as
Left visual point image { the L of definition distortion
Dis(x, y) } in l 8 * 8 sub-pieces be the current second sub-piece, be designated as
Wherein,
(x
2, y
2) the expression current first sub-piece
With the current second sub-piece
The coordinate position of middle pixel, 1≤x
2≤8,1≤y
2≤8,
Represent the current first sub-piece
Middle coordinate position is (x
2, y
2) the pixel value of pixel,
Represent the current second sub-piece
Middle coordinate position is (x
2, y
2) the pixel value of pixel;
3.-2, calculate the current first sub-piece respectively
With the current second sub-piece
In the Grad of all pixels, for the current first sub-piece
Middle coordinate position is (x
2', y
2') pixel, its Grad is designated as P
o(x
2', y
2'), P
o(x
2', y
2')=| G
Ox(x
2', y
2') |+| G
Oy(x
2', y
2') |, for the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel, its Grad is designated as P
d(x
2', y
2'), P
d(x
2', y
2')=| G
Dx(x
2', y
2') |+| G
Dy(x
2', y
2') |, wherein, 1≤x
2'≤8,1≤y
2'≤8, G
Ox(x
2', y
2') the expression current first sub-piece
Middle coordinate position is (x
2', y
2') the horizontal gradient value of pixel, G
Oy(x
2', y
2') the expression current first sub-piece
Middle coordinate position is (x
2', y
2') the vertical gradient value of pixel, G
Dx(x
2', y
2') the expression current second sub-piece
Middle coordinate position is (x
2', y
2') the horizontal gradient value of pixel, G
Dy(x
2', y
2') the expression current second sub-piece
Middle coordinate position is (x
2', y
2') the vertical gradient value of pixel, " || " is for asking absolute value sign;
3.-3, find out the current first sub-piece
In the maximum of Grad of all pixels, be designated as G
Max, then according to G
MaxCalculate first Grads threshold and second Grads threshold, be designated as T respectively
1And T
2, T
1=0.12 * G
Max, T
2=0.06 * G
Max
3.-4, for the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel, judge P
o(x
2', y
2')>T
1And P
d(x
2', y
2')>T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be strong fringe region, Num
1=Nun
1+ 1, then execution in step 3.-8, otherwise, execution in step 3.-5, wherein, Num
1Initial value be 0;
3.-5, judge P
o(x
2', y
2')>T
1And P
d(x
2', y
2')<=T
1, perhaps P
d(x
2', y
2')>T
1And P
o(x
2', y
2')<=T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be weak fringe region, Num
2=Num
2+ 1, then execution in step 3.-8, otherwise, execution in step 3.-6, wherein, Num
2Initial value be 0;
3.-6, judge P
o(x
2', y
2')<T
2And P
d(x
2', y
2')<T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be smooth region, Num
3=Num
3+ 1, then execution in step 3.-8, otherwise, execution in step 3.-7, wherein, Num
3Initial value be 0;
3.-7, judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be texture region, Num
4=Num
4+ 1, wherein, Num
4Initial value be 0;
3.-8, return step and 3.-4 continue remaining pixel in current first sub-piece
and the current second sub-piece
is handled, 8 * 8 pixels in current first sub-piece
and the current second sub-piece
all dispose;
3.-9, with Num
1, Num
2, Num
3And Num
4In maximum The corresponding area type as the current first sub-piece
With the current second sub-piece
Block type, be designated as p, wherein, { 1,2,3,4}, p=1 represent strong edge block to p ∈, and p=2 representes weak edge block, and p=3 representes smooth block, and p=4 representes texture block;
3.-10, make l "=l+1, l=l ", with undistorted left visual point image { L
Org(x, y) } in the next one 8 * 8 sub-pieces as the current first sub-piece, with the left visual point image { L of distortion
Dis(x, y) } in the next one 8 * 8 sub-pieces as the current second sub-piece, return step and 3.-2 continue to carry out, until undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in
8 * 8 sub-pieces of individual non-overlapping copies all dispose, and obtain undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in the block type of all 8 * 8 sub-pieces, wherein, l " initial value be 0;
3.-11,3.-1 employing and step to 3.-10 identical operations, obtain undistorted right visual point image { R
Org(x, y) } and the right visual point image { R of distortion
Dis(x, y) } in the block type of all 8 * 8 sub-pieces.
Described step detailed process 4. is:
4.-1, the left visual point image { L of calculated distortion
Dis(x, y) } in all block types be the spatial noise intensity that is used to reflect picture quality of 8 * 8 sub-pieces of k, be designated as { fq
k(x
2, y
2), for the left visual point image { L of distortion
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel, use it for the reflection picture quality spatial noise intensity be designated as fq
k(x
2, y
2),
Wherein, k ∈ { p|1≤p≤4}, fq
k(x
2, y
2) expression distortion left visual point image { L
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) the spatial noise intensity that is used to reflect picture quality of pixel, 1≤x
2≤8,1≤y
2≤8, N
kLeft visual point image { the L of expression distortion
Dis(x, y) } in block type be the number of 8 * 8 sub-pieces of k, ST
kFor describing the saturation threshold value of error perception, max () is for getting max function, and min () is for getting minimum value function, (x
3, y
3) expression distortion left visual point image { L
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel at undistorted left visual point image { L
Org(x, y) } or undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } in coordinate position, 1≤x
3≤W, 1≤y
3≤H, L
Org(x
3, y
3) expression { L
Org(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, L
Dis(x
3, y
3) expression { L
Dis(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, J
L(x
3, y
3) expression { J
L(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, " || " is for asking absolute value sign;
4.-2, with the left visual point image { L of distortion
Dis(x, y) } in the spatial noise intensity that is used to reflect picture quality of 8 * 8 sub-pieces of various block types be expressed as { fq with set
k(x
2, y
2) | 1≤k≤4}, then with { fq
k(x
2, y
2) | all elements among 1≤k≤4} is arranged in order and is obtained first characteristic vector, is designated as F
1, wherein, F
1Dimension be 256;
4.-3, to undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in each 8 * 8 sub-piece implement singular value decomposition respectively, obtain undistorted left visual point image { L respectively
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in each self-corresponding singular value vector of each 8 * 8 sub-piece, with undistorted left visual point image { L
Org(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Left visual point image { L with distortion
Dis(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Wherein, the dimension of singular value vector is 8,
4.-4, the left visual point image { L of calculated distortion
Dis(x, y) } in all block types be the space structure intensity that is used to reflect picture quality of 8 * 8 sub-pieces of k, be designated as
Wherein, the left visual point image { L of l ' expression distortion
Dis(x, y) } in block type be that 8 * 8 sub-pieces of k are at undistorted left visual point image { L
Org(x, y) } or undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } in sequence number;
4.-5, with the left visual point image { L of distortion
Dis(x, y) } in the space structure intensity that is used to reflect picture quality of 8 * 8 sub-pieces of various block types be expressed as with set
Then will
In all elements arrange in order and obtain second characteristic vector, be designated as F
2, wherein, F
2Dimension be 32;
4.-6, with the first characteristic vector F
1With the second characteristic vector F
2Form new characteristic vector, as the left visual point image { L of distortion
Dis(x, y) } the characteristic vector that is used to reflect picture quality, be designated as F
L, F
L=[F
1, F
2], wherein, F
LDimension be 288, " [] " is the vector representation symbol, [F
1, F
2] represent the first characteristic vector F
1With the second characteristic vector F
2Couple together and form a new characteristic vector;
4.-7, to the right visual point image { R of distortion
Dis(x, y) } adopt with step 4.-1 to 4.-6 identical operations, obtain the right visual point image { R of distortion
Dis(x, y) } the characteristic vector that is used to reflect picture quality, be designated as F
R, wherein, F
RDimension be 288;
4.-8, to the left visual point image { L of distortion
Dis(x, y) } the characteristic vector F that is used to reflect picture quality
LRight visual point image { R with distortion
Dis(x, y) } the characteristic vector F that is used to reflect picture quality
RCarry out linear weighted function, obtain S
DisThe characteristic vector that is used to reflect picture quality, be designated as F
q, F
q=w
L* F
L+ w
R* F
R, wherein, w
LLeft visual point image { the L of expression distortion
Dis(x, y) } weights proportion, w
RRight visual point image { the R of expression distortion
Dis(x, y) } weights proportion, w
L+ w
R=1.
Described step detailed process 5. is:
5.-1, calculate undistorted left visual point image { L respectively
Org(x, y) } and undistorted right visual point image { R
Org(x, y) } absolute difference image, the left visual point image { L of distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } absolute difference image and undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) } absolute difference image, be designated as { D respectively
Org(x, y) }, { D
Dis(x, y) } and Δ J (x, y) }, D
Org(x, y)=| L
Org(x, y)-R
Org(x, y) |, D
Dis(x, y)=| L
Dis(x, y)-R
Dis(x, y) |, Δ J (x, y)=| J
L(x, y)-J
R(x, y) |, wherein, D
Org(x, y) expression { D
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), D
Dis(x, y) expression { D
Dis(x, y) } in coordinate position be (x, the pixel value of pixel y), Δ J (x, y) the middle coordinate position of expression { Δ J (x, y) } is that (" || " is for asking absolute value sign for x, the pixel value of pixel y);
5.-2,3. identical operations of employing and step, obtain { D respectively
Org(x, y) } and { D
Dis(x, y) } in the block type of each 8 * 8 sub-piece;
5.-3, calculate { D
Dis(x, y) } in all block types be the spatial noise intensity that is used to reflect depth perception of 8 * 8 sub-pieces of k, be designated as { fd
k(x
2, y
2), for { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel, use it for the reflection depth perception spatial noise intensity be designated as fd
k(x
2, y
2),
Wherein, fd
k(x
2, y
2) expression { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) the spatial noise intensity that is used to reflect depth perception of pixel, 1≤x
2≤8,1≤y
2≤8, M
kExpression { D
Dis(x, y) } in block type be the number of 8 * 8 sub-pieces of k, ST
kFor describing the saturation threshold value of error perception, (x
4, y
4) expression { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel at { D
Org(x, y) } or { Δ J (x, y) } in coordinate position, 1≤x
4≤W, 1≤y
4≤H, D
Org(x
4, y
4) expression { D
Org(x, y) } in coordinate position be (x
4, y
4) the pixel value of pixel, D
Dis(x
4, y
4) expression { D
Dis(x, y) } in coordinate position be (x
4, y
4) the pixel value of pixel, Δ J (x
4, y
4) coordinate position is (x in the expression { Δ J (x, y) }
4, y
4) the pixel value of pixel;
5.-4, with { D
Dis(x, y) } in the spatial noise intensity that is used to reflect depth perception of 8 * 8 sub-pieces of various block types represent { fd with set
k(x
2, y
2) | 1≤k≤4}, then with { fd
k(x
2, y
2) | all elements among 1≤k≤4} is arranged in order and is obtained the 3rd characteristic vector, is designated as F
3, wherein, F
3Dimension be 256;
5.-5, to { D
Org(x, y) } and { D
Dis(x, y) } in each 8 * 8 sub-piece implement singular value decomposition respectively, obtain { D respectively
Org(x, y) } and { D
Dis(x, y) } in each self-corresponding singular value vector of each 8 * 8 sub-piece, with { D
Org(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
With { D
Dis(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Wherein, the dimension of singular value vector is 8,
5.-6, calculate { D
Dis(x, y) } in all block types be the space structure intensity that is used to reflect depth perception of 8 * 8 sub-pieces of k, be designated as
Wherein, l " expression { D
Dis(x, y) } in block type be that 8 * 8 sub-pieces of k are at { D
Org(x, y) } or { Δ J (x, y) } in sequence number;
5.-7, with { D
Dis(x, y) } in the space structure intensity that is used to reflect depth perception of 8 * 8 sub-pieces of various block types be expressed as with set
Then will
In all elements arrange in order and obtain the 4th characteristic vector, be designated as F
4, wherein, F
4Dimension be 32;
5.-8, with the 3rd characteristic vector F
3With the 4th characteristic vector F
4Form new characteristic vector, as S
DisThe characteristic vector that is used to reflect depth perception, be designated as F
s, F
s=[F
3, F
4], wherein, F
sDimension be 288, " [] " is the vector representation symbol, [F
3, F
4] represent the 3rd characteristic vector F
3With the 4th characteristic vector F
4Couple together and form a new characteristic vector.
Described step detailed process 9. is:
9.-1, the stereo-picture of all distortions of same type of distortion in the distortion stereo-picture set is divided into mutually disjoint 5 groups of subclass, select 4 groups of subclass composing training sample datas set wherein arbitrarily, be designated as Ω
q, { X
k, DMOS
k∈ Ω
q, wherein, q representes training sample data set omega
qIn the width of cloth number of stereo-picture of the distortion that comprises, X
kExpression training sample data set omega
qIn the characteristic vector of stereo-picture of k width of cloth distortion, DMOS
kExpression training sample data set omega
qIn the average subjective scoring difference of stereo-picture of k width of cloth distortion, 1≤k≤q;
9.-2, structure X
kRegression function f (X
k),
Wherein, f () is the function representation form, and w is a weight vector, w
TBe the transposed matrix of w, b is a bias term,
Expression training sample data set omega
qIn the characteristic vector X of stereo-picture of k width of cloth distortion
kLinear function,
D (X
k, X
l) be the kernel function in the support vector regression,
X
lBe training sample data set omega
qIn the characteristic vector of stereo-picture of l width of cloth distortion, γ is a nuclear parameter, is used to reflect the scope of importing sample value; The scope of sample value is big more, and the γ value is also just big more, and exp () expression is the exponential function at the end with e; E=2.71828183, " || || for asking the Euclidean distance symbol;
9.-3, adopt support vector regression to training sample data set omega
qIn the characteristic vector of stereo-picture of all distortion train, make that the regression function value that obtains through training is minimum with the error between the average subjective scoring difference, match obtains the weight vector w of optimum
OptBias term b with optimum
Opt, with the weight vector w of optimum
OptBias term b with optimum
OptCombination be designated as (w
Opt, b
Opt),
The weight vector w of the optimum that utilization obtains
OptBias term b with optimum
OptStructure support vector regression training pattern is designated as
Wherein, ψ representes training sample data set omega
qIn the set of combination of the characteristic vector of stereo-picture of all distortion all weight vector of training and bias term,
Expression minimizes probability density function, X
InpExpress support for the input vector of vector regression training pattern, (w
Opt)
TBe w
OptTransposed matrix,
Express support for the input vector X of vector regression training pattern
InpLinear function;
9.-4, according to the support vector regression training pattern; Stereo-picture to remaining the every width of cloth distortion in 1 group of subclass is tested; Prediction obtains the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion in this group subclass; Evaluating objective quality predicted value for the stereo-picture of j width of cloth distortion in this group subclass is designated as Q with it
j, Q
j=f (X
j),
Wherein, X
jThe characteristic vector of representing the stereo-picture of j width of cloth distortion in this group subclass,
The linear function of representing the stereo-picture of j width of cloth distortion in this group subclass;
9.-5, according to step 9.-1 to 9.-4 process; Respectively the stereo-picture of all distortions of different type of distortion in the set of distortion stereo-picture is trained, obtain the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion in the set of distortion stereo-picture.
Described step 4. with step 8. in the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the JPEG compression artefacts, get w
L=0.50, w
R=0.50; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the JPEG2000 compression artefacts, get w
L=0.15, w
R=0.85; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the Gaussian Blur distortion, get w
L=0.10, w
R=0.90; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the white noise distortion, get w
L=0.20, w
R=0.80; Calculating the H.264 characteristic vector process that is used for reflecting picture quality of the stereo-picture of coding distortion, get w
L=0.10, w
R=0.90.
Compared with prior art, the invention has the advantages that:
1) the inventive method considers that the perception of zones of different stereo has different responses; Stereo-picture is divided into strong edge block, weak edge block, flat block and texture block also to be estimated respectively; Simultaneously picture quality and depth perception information are attached in the evaluation procedure, make evaluation result feel to meet the human visual system more.
2) the inventive method obtains minimum discernable distorted image according to the visual characteristic of human eye; And extract the characteristic information of zones of different piece respectively and form the characteristic vector of stereo-picture through computer memory noise intensity and space structure intensity; The characteristic vector information of the stereo-picture that obtains has stronger stability and can reflect the mass change situation of stereo-picture preferably, has improved the correlation of objective evaluation result and subjective perception.
Embodiment
Embodiment describes in further detail the present invention below in conjunction with accompanying drawing.
A kind of stereo image quality method for objectively evaluating that the present invention proposes based on visually-perceptible, it realizes that totally block diagram is as shown in Figure 1, it mainly may further comprise the steps:
1. make S
OrgUndistorted stereo-picture for original makes S
DisFor the stereo-picture of distortion to be evaluated, with S
OrgLeft visual point image be designated as { L
Org(x, y) }, with S
OrgRight visual point image be designated as { R
Org(x, y) }, with S
DisLeft visual point image be designated as { L
Dis(x, y) }, with S
DisRight visual point image be designated as { R
Dis(x, y) }, wherein, (x, the y) coordinate position of pixel in left visual point image of expression and the right visual point image, 1≤x≤W, 1≤y≤H, W represent the width of left visual point image and right visual point image, H representes the height of left visual point image and right visual point image, L
Org(x, y) expression S
OrgLeft visual point image { L
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), R
Org(x, y) expression S
OrgRight visual point image { R
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), L
Dis(x, y) expression S
DisLeft visual point image { L
Dis(x, y) } in coordinate position be (x, the pixel value of pixel y), R
Dis(x, y) expression S
DisRight visual point image { R
Dis(x, y) } in coordinate position be (x, the pixel value of pixel y).
2. human visual system (HVS) characteristic shows; Human eye is non to changing less attribute or noise in the image; Only if the change intensity of this attribute or noise surpasses a certain threshold value, this threshold value be exactly minimum discernable distortion (Just noticeable distortion, JND).And the visual masking effect of human eye is a kind of local effect, receives the influence of factors such as background illuminance, texture complexity, and background is bright more, and texture is more complicated, and boundary value is just high more.Therefore the present invention utilizes the visual masking effect of human vision to background illumination and texture, extracts undistorted left visual point image { L respectively
Org(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image, with undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image be designated as { J
L(x, y) }, with undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image be designated as { J
R(x, y) }, wherein, J
L(x, y) expression { J
L(x, y) } in coordinate position be (x, the pixel value of pixel y), J
R(x, y) expression { J
R(x, y) } in coordinate position be (x, the pixel value of pixel y).
In this specific embodiment, step detailed process 2. is:
2.-1, calculate undistorted left visual point image { L
Org(x, y) } the visual threshold value set of visual masking effect of background illumination, be designated as { T
t(x, y) },
Wherein, T
l(x, y) the undistorted left visual point image { L of expression
Org(x, y) } in coordinate position be (x, the visual threshold value of the visual masking effect of the background illumination of pixel y),
Represent undistorted left visual point image { L
Org(x, y) } in be that (x, pixel y) they are the average brightness of all pixels in 5 * 5 windows at center with coordinate position; In actual process; Also can adopt other big or small window, but through a large amount of experiments, the result can obtain best effect when showing the window that adopts 5 * 5 sizes.
2.-2, calculate undistorted left visual point image { L
Org(x, y) } the visual threshold value set of visual masking effect of texture, be designated as { T
t(x, y) }, T
t(x, y)=η * G (x, y) * W
e(x, y), wherein, T
t(x, y) the undistorted left visual point image { L of expression
Org(x, y) } in coordinate position be (η is the controlling elements greater than 0 for x, the visual threshold value of the visual masking effect of the texture of pixel y), in the present embodiment, η=0.05, (x y) representes undistorted left visual point image { L G
Org(x, y) } in coordinate position be that (x, pixel y) carry out the maximum weighted mean value that directed high-pass filtering obtains, W
e(x, y) expression is to undistorted left visual point image { L
Org(x, y) } edge image in coordinate position be that (x, pixel y) carry out the edge weighted value that Gauss's LPF obtains.
2.-3, to undistorted left visual point image { L
Org(x, y) } the visual threshold value set { T of visual masking effect of background illumination
l(x, y) } and the visual threshold value set { T of the visual masking effect of texture
t(x, y) } merge, obtain undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image, be designated as { J
L(x, y) }, J
L(x, y)=T
l(x, y)+T
t(x, y)-C
L, t* min{T
l(x, y), T
t(x, y) }, wherein, C
L, tThe parameter of the visual masking effect eclipse effect of expression control background illumination and texture, 0<C
L, t<1, in the present embodiment, C
L, t=0.5, min{} is for getting minimum value function.
2.-4,2.-1 employing and step to 2.-3 identical operations, obtain undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image, be designated as { J
R(x, y) }.
3. because the human visual system has different susceptibilitys to edge of image, texture, smooth region, the responsiveness of zones of different stereo perception also should be different, therefore, in the stereo image quality evaluation, should consider the contribution of zones of different to estimating respectively.The present invention obtains undistorted left visual point image { L respectively through regional detection algorithm
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) }, undistorted right visual point image { R
Org(x, y) } and the right visual point image { R of distortion
Dis(x, y) } in the block type of each 8 * 8 sub-piece, be designated as p, wherein, { 1,2,3,4}, p=1 represent strong edge block to p ∈, and p=2 representes weak edge block, and p=3 representes smooth block, and p=4 representes texture block.
In this specific embodiment, the detailed process of the regional detection algorithm of step in 3. is:
3.-1, respectively with undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } be divided into
8 * 8 sub-pieces of individual non-overlapping copies define undistorted left visual point image { L
Org(x, y) } in l 8 * 8 sub-pieces be the current first sub-piece, be designated as
Left visual point image { the L of definition distortion
Dis(x, y) } in l 8 * 8 sub-pieces be the current second sub-piece, be designated as
Wherein,
(x
2, y
2) the expression current first sub-piece
With the current second sub-piece
The coordinate position of middle pixel, 1≤x
2≤8,1≤y
2≤8,
Represent the current first sub-piece
Middle coordinate position is (x
2, y
2) the pixel value of pixel,
Represent the current second sub-piece
Middle coordinate position is (x
2, y
2) the pixel value of pixel.
3.-2, calculate the current first sub-piece respectively
With the current second sub-piece
In the Grad of all pixels, for the current first sub-piece
Middle coordinate position is (x
2', y
2') pixel, its Grad is designated as P
o(x
2', y
2'), P
o(x
2', y
2')=| G
Ox(x
2', y
2') |+| G
Oy(x
2', y
2') |, for the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel, its Grad is designated as P
d(x
2', y
2'), P
d(x
2', y
2')=| G
Dx(x
2', y
2') |+G
Dy(x
2', y
2') |, wherein, 1≤x
2'≤8,1≤y
2'≤8, G
Ox(x
2', y
2') the expression current first sub-piece
Middle coordinate position is (x
2', y
2') the horizontal gradient value of pixel, G
Oy(x
2', y
2') the expression current first sub-piece
Middle coordinate position is (x
2', y
2') the vertical gradient value of pixel, G
Dx(x
2', y
2') the expression current second sub-piece
Middle coordinate position is (x
2', y
2') the horizontal gradient value of pixel, G
Dy(x
2', y
2') the expression current second sub-piece
Middle coordinate position is (x
2', y
2') the vertical gradient value of pixel, " || " is for asking absolute value sign.
3.-3, find out the current first sub-piece
In the maximum of Grad of all pixels, be designated as G
Max, then according to G
MaxCalculate first Grads threshold and second Grads threshold, be designated as T respectively
1And T
2, T
1=0.12 * G
Max, T
2=0.06 * G
Max
3.-4, for the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel, judge P
o(x
2', y
2')>T
1And P
d(x
2', y
2')>T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be strong fringe region, Num
1=Num
1+ 1, then execution in step 3.-8, otherwise, execution in step 3.-5, wherein, Num
1Initial value be 0.
3.-5, judge P
o(x
2', y
2')>T
1And P
d(x
2', y
2')<=T
1, perhaps P
d(x
2', y
2')>T
1And P
o(x
2', y
2')<=T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be weak fringe region, Num
2=Num
2+ 1, then execution in step 3.-8, otherwise, execution in step 3.-6, wherein, Num
2Initial value be 0.
3.-6, judge P
o(x
2', y
2')<T
2And P
d(x
2', y
2')<T
1Whether set up, if then judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be smooth region, Num
3=Num
3+ 1, then execution in step 3.-8, otherwise, execution in step 3.-7, wherein, Num
3Initial value be 0.
3.-7, judge the current first sub-piece
Middle coordinate position is (x
2', y
2') the pixel and the current second sub-piece
Middle coordinate position is (x
2', y
2') pixel be texture region, Num
4=Num
4+ 1, wherein, Num
4Initial value be 0.
3.-8, return step and 3.-4 continue remaining pixel in current first sub-piece
and the current second sub-piece
is handled, 8 * 8 pixels in current first sub-piece
and the current second sub-piece
all dispose.
3.-9, with Num
1, Num
2, Num
3And Num
4In maximum The corresponding area type as the current first sub-piece
With the current second sub-piece
Block type, be designated as p, wherein, { 1,2,3,4}, p=1 represent strong edge block to p ∈, and p=2 representes weak edge block, and p=3 representes smooth block, and p=4 representes texture block.
3.-10, make l "=l+1, l=l ", with undistorted left visual point image { L
Org(x, y) } in the next one 8 * 8 sub-pieces as the current first sub-piece, with the left visual point image { L of distortion
Dis(x, y) } in the next one 8 * 8 sub-pieces as the current second sub-piece, return step and 3.-2 continue to carry out, until undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in
8 * 8 sub-pieces of individual non-overlapping copies all dispose, and obtain undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in the block type of all 8 * 8 sub-pieces, wherein, l " initial value be 0.
3.-11,3.-1 employing and step to 3.-10 identical operations, obtain undistorted right visual point image { R
Org(x, y) } and the right visual point image { R of distortion
Dis(x, y) } in the block type of all 8 * 8 sub-pieces.
4. because the quality of stereo image quality is directly relevant with left and right sides view-point image quality; In image quality evaluation, introduce the correlation that vision perception characteristics such as visual sensitivity, multichannel characteristic, masking effect can improve evaluation model and subjective scoring; Consider distortion perceptibility and perception saturation phenomenon, with the discernable distorted image of minimum as vision perception characteristic.The present invention is according to undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) }, through the left visual point image { L of calculated distortion
Dis(x, y) } in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects picture quality and the space structure intensity that is used to reflect picture quality and the right visual point image { R of distortion
Dis(x, y) } in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects picture quality and the space structure intensity that is used to reflect picture quality, obtain the left visual point image { L of distortion respectively
Dis(x, y) } be used to reflect the characteristic vector of picture quality and the right visual point image { R of distortion
Dos(x, y) } the characteristic vector that is used to reflect picture quality, again to the left visual point image { L of distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } be used to reflect that the characteristic vector of picture quality carries out linear weighted function, obtains S
DisThe characteristic vector that is used to reflect picture quality, be designated as F
q
In this specific embodiment, step detailed process 4. is:
4.-1, the left visual point image { L of calculated distortion
Dis(x, y) } in all block types be the spatial noise intensity that is used to reflect picture quality of 8 * 8 sub-pieces of k, be designated as { fq
k(x
2, y
2), for the left visual point image { L of distortion
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel, use it for the reflection picture quality spatial noise intensity be designated as fq
k(x
2, y
2),
Wherein, k ∈ { p|1≤p≤4}, fq
k(x
2, y
2) expression distortion left visual point image { L
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) the spatial noise intensity that is used to reflect picture quality of pixel, 1≤x
2≤8,1≤y
2≤8, N
kLeft visual point image { the L of expression distortion
Dis(x, y) } in block type be the number of 8 * 8 sub-pieces of k, ST
kFor describing the saturation threshold value of error perception, in the present embodiment, ST
k=30, max () is for getting max function, and min () is for getting minimum value function, (x
3, y
3) expression distortion left visual point image { L
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel at undistorted left visual point image { L
Org(x, y) } or undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } in coordinate position, 1≤x
3≤W, 1≤y
3≤H, L
Org(x
3, y
3) expression { L
Org(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, L
Dis(x
3, y
3) expression { L
Dis(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, J
L(x
3, y
3) expression { J
L(x, y) } in coordinate position be (x
3, y
3) the pixel value of pixel, " || " is for asking absolute value sign.
4.-2, with the left visual point image { L of distortion
Dis(x, y) } in the spatial noise intensity that is used to reflect picture quality of 8 * 8 sub-pieces of various block types be expressed as { fq with set
k(x
2, y
2) | 1≤k≤4}, then with { fq
k(x
2, y
2) | all elements among 1≤k≤4} is arranged in order and is obtained first characteristic vector, is designated as F
1, wherein, F
1Dimension be 256.
4.-3, to undistorted left visual point image { L
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in each 8 * 8 sub-piece implement singular value decomposition respectively, obtain undistorted left visual point image { L respectively
Org(x, y) } and the left visual point image { L of distortion
Dis(x, y) } in each self-corresponding singular value vector of each 8 * 8 sub-piece, with undistorted left visual point image { L
Org(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Left visual point image { L with distortion
Dis(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Wherein, the dimension of singular value vector is 8,
4.-4, the left visual point image { L of calculated distortion
Dis(x, y) } in all block types be the space structure intensity that is used to reflect picture quality of 8 * 8 sub-pieces of k, be designated as
Wherein, the left visual point image { L of l ' expression distortion
Dis(x, y) } in block type be that 8 * 8 sub-pieces of k are at undistorted left visual point image { L
Org(x, y) } or undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } in sequence number.
4.-5, with the left visual point image { L of distortion
Dis(x, y) } in the space structure intensity that is used to reflect picture quality of 8 * 8 sub-pieces of various block types be expressed as with set
Then will
In all elements arrange in order and obtain second characteristic vector, be designated as F
2, wherein, F
2Dimension be 32.
4.-6, with the first characteristic vector F
1With the second characteristic vector F
2Form new characteristic vector, as the left visual point image { L of distortion
Dis(x, y) } the characteristic vector that is used to reflect picture quality, be designated as F
L, F
L=[F
1, F
2], wherein, F
LDimension be 288, " [] " is the vector representation symbol, [F
1, F
2] represent the first characteristic vector F
1With the second characteristic vector F
2Couple together and form a new characteristic vector.
4.-7, to the right visual point image { R of distortion
Dis(x, y) } adopt with step 4.-1 to 4.-6 identical operations, obtain the right visual point image { R of distortion
Dis(x, y) } the characteristic vector that is used to reflect picture quality, be designated as F
R, wherein, F
RDimension be 288.
4.-8, to the left visual point image { L of distortion
Dis(x, y) } the characteristic vector F that is used to reflect picture quality
LRight visual point image { R with distortion
Dis(x, y) } the characteristic vector F that is used to reflect picture quality
RCarry out linear weighted function, obtain S
DisThe characteristic vector that is used to reflect picture quality, be designated as F
q, F
q=w
L* F
L+ w
R* F
R, wherein, w
LLeft visual point image { the L of expression distortion
Dis(x, y) } weights proportion, w
RRight visual point image { the R of expression distortion
Dis(x, y) } weights proportion, w
L+ w
R=1.
5. existing research shows; When the difference of the absolute difference image of the left and right sides visual point image of the absolute difference image of undistorted left and right sides visual point image and distortion surpasses some threshold values; Then the decline of depth perception is just discovered by human eye easily, therefore, can estimate the depth perception of stereo-picture with the similarity of the absolute difference image of the left and right sides visual point image of the absolute difference image of undistorted left and right sides visual point image and distortion; Absolute difference image is similar more, and depth perception is strong more.Therefore, the present invention is according to undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) }, through the left visual point image { L of calculated distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } absolute difference image in various block types 8 * 8 sub-pieces be used to spatial noise intensity that reflects depth perception and the space structure intensity that is used to reflect depth perception, obtain S
DisThe characteristic vector that is used to reflect depth perception, be designated as F
s
In this specific embodiment, step detailed process 5. is:
5.-1, calculate undistorted left visual point image { L respectively
Org(x, y) } and undistorted right visual point image { R
Org(x, y) } absolute difference image, the left visual point image { L of distortion
Dis(x, y) } and the right visual point image { R of distortion
Dis(x, y) } absolute difference image and undistorted left visual point image { L
Org(x, y) } minimum discernable distorted image { J
L(x, y) } and undistorted right visual point image { R
Org(x, y) } minimum discernable distorted image { J
R(x, y) } absolute difference image, be designated as { D respectively
Org(x, y) }, { D
Dis(x, y) } and Δ J (x, y) }, D
Org(x, y)=| L
Org(x, y)-R
Org(x, y) |, D
Dis(x, y)=| L
Dis(x, y)-R
Dis(x, y) |, Δ J (x, y)=| J
L(x, y)-J
R(x, y) |, wherein, D
Org(x, y) expression { D
Org(x, y) } in coordinate position be (x, the pixel value of pixel y), D
Dis(x, y) expression { D
Dis(x, y) } in coordinate position be that ((x, y) the middle coordinate position of expression { Δ J (x, y) } is that (" || " is for asking absolute value sign for x, the pixel value of pixel y) to Δ J for x, the pixel value of pixel y).
5.-2,3. identical operations of employing and step, obtain { D respectively
Org(x, y) } and { D
Dis(x, y) } in the block type of each 8 * 8 sub-piece.
5.-3, calculate { D
Dis(x, y) } in all block types be the spatial noise intensity that is used to reflect depth perception of 8 * 8 sub-pieces of k, be designated as { fd
k(x
2, y
2), for { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel, use it for the reflection depth perception spatial noise intensity be designated as fd
k(x
2, y
2),
Wherein, fd
k(x
2, y
2) expression { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) the spatial noise intensity that is used to reflect depth perception of pixel, 1≤x
2≤8,1≤y
2≤8, M
kExpression { D
Dis(x, y) } in block type be the number of 8 * 8 sub-pieces of k, ST
kFor describing the saturation threshold value of error perception, (x
4, y
4) expression { D
Dis(x, y) } in block type be that coordinate position is (x in 8 * 8 sub-pieces of k
2, y
2) pixel at { D
Org(x, y) } or { Δ J (x, y) } in coordinate position, 1≤x
4≤W, 1≤y
4≤H, D
Org(x
4, y
4) expression { D
Org(x, y) } in coordinate position be (x
4, y
4) the pixel value of pixel, D
Dis(x
4, y
4) expression { D
Dis(x, y) } in coordinate position be (x
4, y
4) the pixel value of pixel, Δ J (x
4, y
4) coordinate position is (x in the expression { Δ J (x, y) }
4, y
4) the pixel value of pixel.
5.-4, with { D
Dis(x, y) } in the spatial noise intensity that is used to reflect depth perception of 8 * 8 sub-pieces of various block types represent { fd with set
k(x
2, y
2) | 1≤k≤4}, then with { fd
k(x
2, y
2) | all elements among 1≤k≤4} is arranged in order and is obtained the 3rd characteristic vector, is designated as F
3, wherein, F
3Dimension be 256.
5.-5, to { D
Org(x, y) } and { D
Dis(x, y) } in each 8 * 8 sub-piece implement singular value decomposition respectively, obtain { D respectively
Org(x, y) } and { D
Dis(x, y) } in each self-corresponding singular value vector of each 8 * 8 sub-piece, with { D
Org(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
With { D
Dis(x, y) } in the singular value vector of l 8 * 8 sub-pieces be designated as
Wherein, the dimension of singular value vector is 8,
5.-6, calculate { D
Dis(x, y) } in all block types be the space structure intensity that is used to reflect depth perception of 8 * 8 sub-pieces of k, be designated as
Wherein, l " expression { D
Dis(x, y) } in block type be that 8 * 8 sub-pieces of k are at { D
Org(x, y) } or { Δ J (x, y) } in sequence number.
5.-7, with { D
Dis(x, y) } in the space structure intensity that is used to reflect depth perception of 8 * 8 sub-pieces of various block types be expressed as with set
Then will
In all elements arrange in order and obtain the 4th characteristic vector, be designated as F
4, wherein, F
4Dimension be 32.
5.-8, with the 3rd characteristic vector F
3With the 4th characteristic vector F
4Form new characteristic vector, as S
DisThe characteristic vector that is used to reflect depth perception, be designated as F
s, F
s=[F
3, F
4], wherein, F
sDimension be 288, " [] " is the vector representation symbol, [F
3, F
4] represent the 3rd characteristic vector F
3With the 4th characteristic vector F
4Couple together and form a new characteristic vector.
6. with S
DisThe characteristic vector F that is used to reflect picture quality
qWith the characteristic vector F that is used to reflect depth perception
sForm new characteristic vector, as S
DisCharacteristic vector, be designated as X, X=[F
q, F
s], " [] " is the vector representation symbol, [F
q, F
s] represent characteristic vector F
qWith characteristic vector F
sCouple together and form a new characteristic vector.
7. adopt n undistorted stereo-picture, set up its distortion stereo-picture set under the different distortion levels of different type of distortion, this distortion stereo-picture set comprises the stereo-picture of several distortions; Utilize the subjective quality evaluation method to obtain the average subjective scoring difference of the stereo-picture of every width of cloth distortion in the set of distortion stereo-picture respectively, be designated as DMOS, DMOS=100-MOS; Wherein, MOS representes the subjective scoring average, DMOS ∈ [0; 100], n >=1.
In the present embodiment; Because the stereo-picture of test obtains through H.264 encoding; Therefore the type of distortion of training sample and test sample book should be consistent in support vector regression; Utilize as stereo-picture that Fig. 2 a and stereo-picture, Figure 13 a and Figure 13 b that stereo-picture, Figure 12 a and Figure 12 b that stereo-picture, Figure 11 a and Figure 11 b that stereo-picture, Figure 10 a and Figure 10 b that stereo-picture, Fig. 9 a and Fig. 9 b that stereo-picture, Fig. 8 a and Fig. 8 b that stereo-picture, Fig. 7 a and Fig. 7 b that stereo-picture, Fig. 6 a and Fig. 6 b that stereo-picture, Fig. 5 a and Fig. 5 b that stereo-picture, Fig. 4 a and Fig. 4 b that stereo-picture, Fig. 3 a and Fig. 3 b that Fig. 2 b constitutes constitute constitute constitute constitute constitute constitute constitute constitute constitute constitute constitute the undistorted stereo-picture of totally 12 width of cloth (n=12) set up its distortion stereo-picture under the different distortion levels of coding distortion type H.264 and gathered, the stereo-picture of distortion has 72 width of cloth in this distortion stereo-picture is gathered.
8. adopt and calculating S
DisThe identical method of characteristic vector X, the characteristic vector of the stereo-picture of every width of cloth distortion in the set of calculated distortion stereo-picture respectively, the characteristic vector for the stereo-picture of i width of cloth distortion in the set of distortion stereo-picture is designated as X with it
i, wherein, 1≤i≤n ', the width of cloth number of the stereo-picture of the distortion that comprises in the set of n ' expression distortion stereo-picture.
In this specific embodiment; According to the stereoscopic vision masking effect inconsistent characteristic of human eye to different type of distortion; Left visual point image to the stereo-picture of different type of distortion is provided with different weights proportion with right visual point image; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the JPEG compression artefacts, get w
L=0.50, w
R=0.50; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the JPEG2000 compression artefacts, get w
L=0.15, w
R=0.85; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the Gaussian Blur distortion, get w
L=0.10, w
R=0.90; In the characteristic vector process that is used for reflecting picture quality of the stereo-picture that calculates the white noise distortion, get w
L=0.20, w
R=0.80; Calculating the H.264 characteristic vector process that is used for reflecting picture quality of the stereo-picture of coding distortion, get w
L=0.10, w
R=0.90.
9. because the characteristic vector of the stereo-picture of distortion is the higher dimensional space vector; Need in higher dimensional space, construct linear decision function and realize the non-linear decision function in the former space; (Support Vector Regression SVR) is the method for the non-linear higher dimensional space conversion of a kind of reasonable realization to support vector regression.Therefore the inventive method adopts support vector regression that the characteristic vector of the stereo-picture of all distortions of identical type of distortion in the set of distortion stereo-picture is trained; And the support vector regression training pattern of utilizing training to obtain is tested the stereo-picture of every width of cloth distortion of same type of distortion; Calculate the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion of identical type of distortion in the set of distortion stereo-picture; Evaluating objective quality predicted value for the stereo-picture of i width of cloth distortion in the set of distortion stereo-picture is designated as Q with it
i, Q
i=f (X
i), f () is the function representation form, Q
i=f (X
i) expression Q
iBe X
iFunction, wherein, 1≤i≤n ', the width of cloth number of the stereo-picture of the distortion that comprises in the n ' expression distortion stereo-picture set.
In this specific embodiment, step detailed process 9. is:
9.-1, the stereo-picture of all distortions of same type of distortion in the distortion stereo-picture set is divided into mutually disjoint 5 groups of subclass, select 4 groups of subclass composing training sample datas set wherein arbitrarily, be designated as Ω
q, { X
k, DMOS
k∈ Ω
q, wherein, q representes training sample data set omega
qIn the width of cloth number of stereo-picture of the distortion that comprises, X
kExpression training sample data set omega
qIn the characteristic vector of stereo-picture of k width of cloth distortion, DMOS
kExpression training sample data set omega
qIn the average subjective scoring difference of stereo-picture of k width of cloth distortion, 1≤k≤q.
9.-2, structure X
kRegression function f (X
k),
Wherein, f () is the function representation form, and w is a weight vector, w
TBe the transposed matrix of w, b is a bias term,
Expression training sample data set omega
qIn the characteristic vector X of stereo-picture of k width of cloth distortion
kLinear function,
D (X
k, X
l) be the kernel function in the support vector regression,
X
lBe training sample data set omega
qIn the characteristic vector of stereo-picture of l width of cloth distortion, γ is a nuclear parameter, is used to reflect the scope of importing sample value; The scope of sample value is big more, and the γ value is also just big more, and exp () expression is the exponential function at the end with e; E=2.71828183, " || || for asking the Euclidean distance symbol.
In the present embodiment, JPEG compression artefacts, JPEG 2000 compression artefacts, Gaussian Blur distortion, white noise distortion and H.264 the γ value of coding distortion get 42,52,54,130 and 116 respectively.
9.-3, adopt support vector regression to training sample data set omega
qIn the characteristic vector of stereo-picture of all distortion train, make that the regression function value that obtains through training is minimum with the error between the average subjective scoring difference, match obtains the weight vector w of optimum
OptBias term b with optimum
Opt, with the weight vector w of optimum
OptBias term b with optimum
OptCombination be designated as (w
Opt, b
Opt),
The weight vector w of the optimum that utilization obtains
OptBias term b with optimum
OptStructure support vector regression training pattern is designated as
Wherein, ψ representes training sample data set omega
qIn the set of combination of the characteristic vector of stereo-picture of all distortion all weight vector of training and bias term,
Expression minimizes probability density function, X
InpExpress support for the input vector of vector regression training pattern, (w
Opt)
TBe w
OptTransposed matrix,
Express support for the input vector X of vector regression training pattern
InpLinear function.
9.-4, according to the support vector regression training pattern; Stereo-picture to remaining the every width of cloth distortion in 1 group of subclass is tested; Prediction obtains the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion in this group subclass; Evaluating objective quality predicted value for the stereo-picture of j width of cloth distortion in this group subclass is designated as Q with it
j, Q
j=f (X
j),
Wherein, X
jThe characteristic vector of representing the stereo-picture of j width of cloth distortion in this group subclass,
The linear function of representing the stereo-picture of j width of cloth distortion in this group subclass.
9.-5, according to step 9.-1 to 9.-4 process; Respectively the stereo-picture of all distortions of different type of distortion in the set of distortion stereo-picture is trained, obtain the evaluating objective quality predicted value of the stereo-picture of every width of cloth distortion in the set of distortion stereo-picture.
Adopt 12 undistorted stereo-pictures shown in Fig. 2 a to Figure 13 b to analyze the objective image quality evaluation predicted value and the average correlation between the subjective scoring difference of the stereo-picture of the distortion that present embodiment obtains at the stereo-picture of in various degree JPEG compression, JPEG2000 compression, Gaussian Blur, white noise and H.264 312 width of cloth distortions under the coding distortion situation.Here; 2 objective parameters commonly used that utilize the evaluate image quality evaluating method are as evaluation index; Be under the nonlinear regression condition Pearson correlation coefficient (Correlation Coefficient, CC), the Spearman coefficient correlation (Rank-Order Correlation Coefficient, ROCC); The stereo-picture of CC reflection distortion is estimated the accuracy of objective models, and ROCC reflects its monotonicity.The objective image evaluation quality predicted value of the stereo-picture of the distortion that will calculate by present embodiment is done four parameter L ogistic function nonlinear fittings, and the high more explanation method for objectively evaluating of CC and ROCC value is good more with average subjective scoring difference correlation.Table 1 has been listed the image quality of stereoscopic images of the distortion that present embodiment obtains and has been estimated the correlation between predicted value and the subjective scoring; Can know from the data that table 1 is listed; Correlation between the final objective image quality evaluation predicted value of the stereo-picture of the distortion that obtains by present embodiment and the average subjective scoring difference is very high; The result who shows objective evaluation result and human eye subjective perception is more consistent, is enough to explain the validity of the inventive method.
Table 2 has provided and has adopted the image quality of stereoscopic images of the distortion that the different characteristic vector obtains to estimate the correlation between predicted value and the subjective scoring; From table 2, can find out; Only adopt evaluation predicted value single or that two characteristic vectors obtain all and between the subjective scoring all to have bigger correlation; The feature extracting method that the inventive method is described is effective; And in conjunction with the characteristic vector of reflection picture quality and depth perception, the evaluation predicted value and the correlation between the subjective scoring that obtain are stronger, are enough to explain that this method is effective.
The image quality of stereoscopic images of the distortion that table 1 present embodiment obtains is estimated the correlation between predicted value and the subjective scoring
The image quality of stereoscopic images of the distortion that table 2 employing different characteristic vector obtains is estimated the correlation between predicted value and the subjective scoring