Disclosure of Invention
The technical problem to be solved by the invention is to provide an objective evaluation method for quality of an asymmetric multi-distortion stereo image, which can effectively improve the correlation between objective evaluation results and subjective perception, has low calculation complexity and does not need to predict subjective evaluation values of tested stereo images.
The technical scheme adopted by the invention for solving the technical problems is as follows: an objective evaluation method for quality of asymmetric multi-distortion stereo images is characterized by comprising a training stage and a testing stage;
the specific steps of the training phase process are as follows:
① _1, selecting N original undistorted stereo images with width W and height H, then respectively performing JPEG distortion, Gaussian blur distortion and Gaussian white noise distortion of L different distortion intensities on each original undistorted stereo image to obtain JPEG distorted stereo images of L distortion intensities, Gaussian blur distorted stereo images of L distortion intensities and Gaussian white noise distorted stereo images of L distortion intensities corresponding to each original undistorted stereo image, and then forming a first training image set by all original undistorted stereo images and JPEG distorted stereo images of L distortion intensities corresponding to each original undistorted stereo image, and recording the first training image set as a first training image set
And all original undistorted stereo images and L corresponding Gaussian blur distorted stereo images with distortion intensity form a second training image set which is recorded as
All original undistorted stereo images and L distortion intensity Gaussian white noise distorted stereo images corresponding to the original undistorted stereo images form a third training image set, and the third training image set is recorded as
Wherein N is>1,L>1,
To represent
And
the u-th original undistorted stereo image in (a),
to represent
The distorted stereo image of the v-th distortion intensity corresponding to the original undistorted stereo image in (1),
to represent
The distorted stereo image of the v-th distortion intensity corresponding to the original undistorted stereo image in (1),
to represent
The distortion stereo image with the v distortion intensity corresponding to the u original distortion-free stereo image;
① _2, respectively obtaining by 6 different full reference image quality evaluation methods
And
objective evaluation prediction value of each distorted stereo image; then will be
The 6 objective evaluation predicted values of each distorted three-dimensional image form the image quality vector of the distorted three-dimensional image in sequence, and the image quality vector of the distorted three-dimensional image is obtained
The 6 objective evaluation predicted values of each distorted three-dimensional image form the image quality vector of the distorted three-dimensional image in sequence, and the image quality vector of the distorted three-dimensional image is obtained
The 6 objective evaluation predicted values of each distorted three-dimensional image form an image quality vector of the distorted three-dimensional image in sequence;
① _3, will
The image quality vectors and the average subjective score difference value of all the distorted stereo images form a first training sample data set; then, a support vector regression is adopted as a machine learning method to train all image quality vectors in the first training sample data set, so that regression function values and subjects obtained through trainingThe error between the quality recommendation values is minimum, and the optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a first quality prediction model, denoted as g
1(y
1),
Wherein, g
1() In the form of a function, y
1For representing an image quality vector, and as an input vector to a first quality prediction model,
is composed of
The transpose of (a) is performed,
is y
1A linear function of (a);
also, will
The image quality vectors and the average subjective score difference values of all the distorted stereo images form a second training sample data set; then, a support vector regression is adopted as a machine learning method to train all image quality vectors in the second training sample data set, so that the error between a regression function value obtained through training and a subjective quality recommended value is minimum, and an optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a second quality prediction model, denoted as g
2(y
2),
Wherein, g
2() In the form of a function, y
2For representing an image quality vector, and as an input vector to a second quality prediction model,
is composed of
The transpose of (a) is performed,
is y
2A linear function of (a);
also, will
The image quality vectors and the average subjective score difference values of all the distorted stereo images form a third training sample data set; then, a support vector regression is adopted as a machine learning method to train all image quality vectors in the third training sample data set, so that the error between a regression function value obtained through training and a subjective quality recommended value is minimum, and an optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a third quality prediction model, denoted as g
3(y
3),
Wherein, g
3() In the form of a function, y
3For representing an image quality vector, and as an input vector to a third quality prediction model,
is composed of
The transpose of (a) is performed,
is y
3A linear function of (a);
① _4, calculation
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
Also, calculate
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
Also, calculate
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
① _5, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the kth sub-block in all the local phase images in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein, the symbol
Is a sign of a lower rounding operation, k is more than or equal to 1 and less than or equal to M,
and
the dimensions of (A) are all 64 x 1;
also, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
All parts ofThe image feature vector formed by the pixel values of all the pixel points in the kth sub-block in the phase image in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein the content of the first and second substances,
and
the dimensions of (A) are all 64 x 1;
also, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the kth sub-block in all the local phase images in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein the content of the first and second substances,
and
the dimensions of (A) are all 64 x 1;
① _6, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other; then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
1,k(ii) a Then will be
All of the distortions inThe set of image quality vectors of all the subblocks in the stereoscopic image is denoted by { y }
1,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
1,kHas a dimension of 6 × 1;
also, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other; then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
2,k(ii) a Then will be
The set of image quality vectors of all the subblocks in all the distorted stereoscopic images in (1) is denoted by { y }
2,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
2,kHas a dimension of 6 × 1;
also, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other;then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
3k(ii) a Then will be
The set of image quality vectors of all the subblocks in all the distorted stereoscopic images in (1) is denoted by { y }
3,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
3,kHas a dimension of 6 × 1;
① _7, pair of bags by K-SVD method
And
performing joint dictionary training operation on the formed set to obtain a structure
Respective image feature dictionary table and image quality dictionary table, corresponding
And
wherein the content of the first and second substances,
and
the dimensions of (a) are all 64 x K,
and
the dimensions of the dictionary are 6 XK, K represents the number of the set dictionaries, and K is more than or equal to 1;
also, the K-SVD method is adopted to pair
And
performing joint dictionary training operation on the formed set to obtain a structure
And
respective image feature dictionary table and image quality dictionary table, corresponding
And
wherein the content of the first and second substances,
and
the dimensions of (a) are all 64 x K,
and
the dimensions of (A) are all 6 XK;
the specific steps of the test phase process are as follows:
② _1, for any test stereo image S with width W' and height HtestWill StestIs recorded as LtestWill StestIs recorded as Rtest(ii) a Wherein W 'is the same as or different from W, and H' is the same as or different from H;
② _2, obtaining S in the same operation as step ① _4
test、L
testAnd R
testRespective local phase image and local amplitude image, and converting L into L
testThe local phase image and the local amplitude image are associated as
And
r is to be
testThe local phase image and the local amplitude image are associated as
And
② _3, will
And
are respectively divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Will be provided with
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Will be provided with
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Then will be
The set of image feature vectors of all sub-blocks in (1) is denoted as
And will be
The set of image feature vectors of all sub-blocks in (1) is denoted as
Will be provided with
The set of image feature vectors of all sub-blocks in (1) is denoted as
Will be provided with
The set of image feature vectors of all sub-blocks in (1) is denoted as
Wherein the content of the first and second substances,
and
the dimensions of (A) are all 64 x 1;
② _4 constructed according to the procedure in the training phase
Separately optimized reconstruction
And
a first sparse coefficient matrix for each respective image feature vector
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
also, in the same manner as above,constructed according to a process during a training phase
Separately optimized reconstruction
And
a second sparse coefficient matrix for each respective image feature vector
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
also constructed from the process during the training phase
Separately optimized reconstruction
And
a third sparse coefficient matrix for each respective image feature vector
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
also constructed from the process during the training phase
Separately optimized reconstruction
And
a first sparse coefficient matrix for each respective image feature vector
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
also constructed from the process during the training phase
Separately optimized reconstruction
And
a second sparse coefficient matrix for each respective image feature vector
Is expressed as a sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
also constructed from the process during the training phase
Separately optimized reconstruction
And
a third sparse coefficient matrix for each respective image feature vector
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtaining;
wherein the content of the first and second substances,
and
all dimensions of (a) are Kx 1, min () is a minimum function, and the symbol "| | | | non-conducting filament
F"is a Flobenius norm-norm symbol for solving the matrix, the symbol" | | | | | | luminance
1"is the 1-norm symbol of the matrix, λ is the Lagrange parameter;
② _5 constructed according to the procedure in the training phase
Estimate separately
And
the first image quality vector of each sub-block in each will
The first image quality vector of the t sub-block of (1) is noted
Will be provided with
The first image quality vector of the t sub-block of (1) is noted
Also constructed from the process during the training phase
Estimate separately
And
the second image quality vector of each sub-block in each will
The second image quality vector of the t sub-block of (1)
Will be provided with
The second image quality vector of the t sub-block of (1)
Also constructed from the process during the training phase
Estimate separately
And
the third image quality vector of each sub-block in each
The third image quality vector of the t sub-block of (1)
Will be provided with
The third image quality vector of the t sub-block of (1)
Also constructed from the process during the training phase
Estimate separately
And
the first image quality vector of each sub-block in each will
The first image quality vector of the t sub-block of (1) is noted
Will be provided with
The first image quality vector of the t sub-block of (1) is noted
Also constructed from the process during the training phase
Estimate separately
And
the second image quality vector of each sub-block in each will
The second image quality vector of the t sub-block of (1)
Will be provided with
The second image quality vector of the t sub-block of (1)
Also constructed from the process during the training phase
Estimate separately
And
the third image quality vector of each sub-block in each
The third image quality vector of the t sub-block of (1)
Will be provided with
The third image quality vector of the t sub-block of (1)
Wherein the content of the first and second substances,
and
the dimensions of (A) are all 6 x 1;
②_6、computing
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein exp () represents an exponential function with a natural base e as a base, and the symbol "| | | | purple
2"is the 2-norm sign of the matrix, η is the control parameter,
is composed of
The input vector of (1);
also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1);
also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1);
also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1);
② _7, calculation
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
L,P,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
R,P,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
,LA,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
R,A,
② _8, according to
And
and Q
L,PAnd Q
R,PCalculating S
testThe predicted value of the objective quality evaluation of the local phase image is recorded as Q
P,Q
P=ω
L,P×Q
L,P+ω
R,P×Q
R,P(ii) a Wherein, ω is
L,PIs Q
L,PThe weight of (a) is calculated,
ω
R,Pis Q
R,PThe weight of (a) is calculated,
symbol'<>"is the inner product symbol, C is the control parameter;
also according to
And
and Q
L,AAnd Q
R,ACalculating S
testThe predicted value of the objective evaluation of the quality of the local amplitude image is marked as Q
A,Q
A=ω
L,A×Q
L,A+ω
R,A×Q
R,A(ii) a Wherein, ω is
L,AIs Q
L,AThe weight of (a) is calculated,
ω
R,Ais Q
R,AThe weight of (a) is calculated,
② _9, according to QPAnd QACalculating StestThe predicted value of the objective evaluation of image quality is expressed as Q, Q ═ ω (ω)P×(QP)n+(1-ωP)×(QA)n)1/n(ii) a Wherein, ω isPAnd n are weighting parameters.
In the step ① _4, the data is sent,
and
the acquisition process comprises the following steps:
① _4a, using Log-Gabor filter pairs
Each pixel point in the image is filtered to obtain
The even symmetric frequency response and the odd symmetric frequency response of each pixel point in different scales and directions will be
The even symmetric frequency response of the pixel point with the middle coordinate position (x, y) in different scales and directions is recorded as e
α,θ(x, y) is
The odd symmetric frequency response of the pixel point with the middle coordinate position (x, y) in different scales and directions is recorded as o
α,θ(x, y), wherein x is more than or equal to 1 and less than or equal to W, y is more than or equal to 1 and less than or equal to H, α represents the scale factor of the Log-Gabor filter,
theta denotes a direction factor of the Log-Gabor filter,
① _4b, calculation
The phase consistency characteristics of each pixel point in different directions are
The phase consistency characteristics of the pixel points with the (x, y) middle coordinate position in different directions are recorded as PC
θ(x,y),
Wherein the content of the first and second substances,
① _4c, according to
The direction corresponding to the maximum phase consistency characteristic of each pixel point in the image is calculated
The local phase characteristic and the local amplitude characteristic of each pixel point in the image; for the
And (3) finding out the maximum phase consistency characteristic of the pixel point with the (x, y) middle coordinate position in the phase consistency characteristics in different directions, finding out the direction corresponding to the maximum phase consistency characteristic, and marking as theta
mAgain according to theta
mCalculating the local phase characteristic and the local amplitude characteristic of the pixel point, and correspondingly marking as
And
wherein the content of the first and second substances,
arctan () is an inverted cosine function,
to represent
The pixel point with the middle coordinate position (x, y) is in the direction theta corresponding to the different scales and the maximum phase consistency characteristics of the pixel point
mThe odd-symmetric frequency response of (a),
to represent
The pixel point with the middle coordinate position (x, y) is in the direction theta corresponding to the different scales and the maximum phase consistency characteristics of the pixel point
mThe even-symmetric frequency response of the frequency domain,
① _4d, according to
The local phase characteristics of all the pixel points in the image are obtained
Local phase image of
Also according to
Obtaining the local amplitude characteristics of all the pixel points in the image
Local amplitude image of
Obtaining according to steps ① _4a through ① _4d
And
in the same manner as the procedure of (1)
And
and
in the step ① _7, the user can,
and
is solved by adopting a K-SVD method
Obtained, where min () is a minimum function, the symbol "| | | | luminance
F"is a Flobenius norm-norm symbol for solving the matrix, the symbol" | | | | | | luminance
1"is to find the 1-norm sign of the matrix, s is more than or equal to 1 and less than or equal to 3,
and
the dimensions of (a) are all 64 x M,
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
The mth first image feature vector in (1),
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
The mth first image feature vector in (1),
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
M-th first image feature vector, Y
1=[y
1,1…y
1,k…y
1,M],Y
2=[y
2,1…y
2,k…y
2,M],Y
3=[y
3,1…y
3,k…y
3,M],Y
1、Y
2And Y
3All dimensions of (a) are 6 XM, y
1,1Is { y
1,k1 st image quality vector, y in |1 ≦ k ≦ M }
1,kIs { y
1,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
1,MIs { y
1,kMth image quality vector in |1 ≦ k ≦ M ≦ y
2,1Is { y
2,k1 st image quality vector, y in |1 ≦ k ≦ M }
2,kIs { y
2,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
2,MIs { y
2,kMth image quality vector in |1 ≦ k ≦ M ≦ y
3,1Is { y
3,k1 st image quality vector, y in |1 ≦ k ≦ M }
3,kIs { y
3,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
3,MIs { y
3,kThe Mth image quality vector in |1 ≦ k ≦ M },
and
each of which represents a sparse matrix and each of which represents,
and
the dimensions of (A) are all K multiplied by M,
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
the dimension of (A) is K x 1, symbol [ "]]"is a vector representation symbol, γ is a weighting parameter, and λ is a lagrangian parameter;
in the step ① _7, the user can,
and
is solved by adopting a K-SVD method
The process for preparing a novel compound of formula (I),
and
all dimension of (A) are 64M,
Is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
and
each of which represents a sparse matrix and each of which represents,
and
the dimensions of (A) are all K multiplied by M,
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
the dimensions of (A) are each K × 1.
Compared with the prior art, the invention has the advantages that:
1) in the training stage, acquiring JPEG (joint photographic experts group) distorted stereo images, Gaussian fuzzy distorted stereo images and Gaussian white noise distorted stereo images with different distortion intensities of undistorted stereo images, respectively constructing three training image sets, and respectively obtaining image feature dictionary tables and image quality dictionary tables of all local phase images and local amplitude images under different distortion types through joint dictionary training; in the testing stage, the image characteristic dictionary table and the image quality dictionary table do not need to be calculated, so that the complex machine learning training process is avoided, the subjective evaluation value of each tested stereo image does not need to be predicted, the calculation complexity is low, and the method is suitable for practical application occasions.
2) In the testing stage, according to the image characteristic dictionary table of the local phase image and the local amplitude image under different distortion types, which is obtained by construction in the training stage, the method obtains the sparse coefficient matrix of each sub-block in the local phase image and the local amplitude image of the tested stereo image through optimization, obtains the image quality vector of each sub-block in the local phase image and the local amplitude image under different distortion types, which are obtained by construction in the training stage, and finally obtains the image quality objective evaluation predicted value of the tested stereo image through multi-distortion fusion, local global fusion, left-right viewpoint fusion and phase amplitude fusion of the sparse coefficient matrix and the image quality vector, and keeps better consistency with the subjective evaluation value.
Detailed Description
The invention is described in further detail below with reference to the accompanying examples.
The overall implementation block diagram of the objective evaluation method for the quality of the asymmetric multi-distortion stereo image provided by the invention is shown in fig. 1, and the method comprises a training stage and a testing stage. The specific steps of the training phase process are as follows:
① _1, selecting N original undistorted stereo images with width W and height H, then respectively performing JPEG distortion, Gaussian blur distortion and Gaussian white noise distortion of L different distortion intensities on each original undistorted stereo image to obtain JPEG distorted stereo images of L distortion intensities, Gaussian blur distorted stereo images of L distortion intensities and Gaussian white noise distorted stereo images of L distortion intensities corresponding to each original undistorted stereo image, and then forming a first training image set by all original undistorted stereo images and JPEG distorted stereo images of L distortion intensities corresponding to each original undistorted stereo image, and recording the first training image set as a first training image set
And all original undistorted stereo images and L corresponding Gaussian blur distorted stereo images with distortion intensity form a second training image set which is recorded as
All original undistorted stereo images and L distortion intensity Gaussian white noise distorted stereo images corresponding to the original undistorted stereo images form a third training image set, and the third training image set is recorded as
Wherein N is>In this example, N is 10, L>1, in this embodiment, L is 3,
to represent
And
the u-th original undistorted stereo image in (a),
to represent
The distorted stereo image of the v-th distortion intensity corresponding to the original undistorted stereo image in (1),
to represent
The distorted stereo image of the v-th distortion intensity corresponding to the original undistorted stereo image in (1),
to represent
The symbol "{ }" is a set representing a symbol of the distorted stereoscopic image of the v-th distortion strength corresponding to the u-th original undistorted stereoscopic image in (a).
In specific implementation, 10 original undistorted stereo images are taken, and each original undistorted stereo image is added with 3 distortion intensity JPEG distortions, 3 distortion intensity Gaussian blur distortions and 3 distortion intensity Gaussian white noise distortions, so that a first training image set consisting of 10 original undistorted stereo images and 30 JPEG distorted stereo images, a second training image set consisting of 10 original undistorted stereo images and 30 Gaussian blur distorted stereo images and a third training image set consisting of 10 original undistorted stereo images and 30 Gaussian white noise distorted stereo images are obtained.
① _2, respectively obtaining by 6 different full reference image quality evaluation methods
And
objective evaluation prediction value of each distorted stereo image; then will be
The 6 objective evaluation predicted values of each distorted three-dimensional image form the image quality vector of the distorted three-dimensional image in sequence, and the image quality vector of the distorted three-dimensional image is obtained
The 6 objective evaluation predicted values of each distorted three-dimensional image form the image quality vector of the distorted three-dimensional image in sequence, and the image quality vector of the distorted three-dimensional image is obtained
The 6 objective evaluation predicted values of each distorted stereo image form the image quality vector of the distorted stereo image in sequence.
In this embodiment, the 6 different full-reference image quality evaluation methods adopted are the known PSNR, MS-SSIM, FSIM, VIF, IW-SSIM, and UQI full-reference image quality evaluation methods, respectively.
① _3, will
The image quality vectors and the average subjective score difference value of all the distorted stereo images form a first training sample data set; then, a method of supporting vector regression is adopted as machine learning, all image quality vectors in the first training sample data set are trained, so that the error between a regression function value obtained through training and a subjective quality recommended value is minimum, and an optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a first quality prediction model, denoted as g
1(y
1),
Wherein, g
1() In the form of a function, y
1For representing an image quality vector, and as an input vector to a first quality prediction model,
is composed of
The transpose of (a) is performed,
is y
1Is a linear function of (a).
Also, will
The image quality vectors and the average subjective score difference values of all the distorted stereo images form a second training sample data set; then, a support vector regression is adopted as a machine learning method to train all image quality vectors in the second training sample data set, so that the error between a regression function value obtained through training and a subjective quality recommended value is minimum, and an optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a second quality prediction model, denoted as g
2(y
2),
Wherein, g
2() In the form of a function, y
2For representing an image quality vector, and as an input vector to a second quality prediction model,
is composed of
The transpose of (a) is performed,
is y
2Is a linear function of (a).
Also, will
The image quality vectors and the average subjective score difference values of all the distorted stereo images form a third training sample data set; then, a support vector regression is adopted as a machine learning method to train all image quality vectors in the third training sample data set, so that the error between a regression function value obtained through training and a subjective quality recommended value is minimum, and an optimal weight vector is obtained through fitting
And an optimal bias term
Then use
And
constructing a third quality prediction model, denoted as g
3(y
3),
Wherein, g
3() In the form of a function, y
3For representing an image quality vector, and as an input vector to a third quality prediction model,
is composed of
The transpose of (a) is performed,
is y
3Is a linear function of (a).
① _4, calculation
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
Also, calculate
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
Also, calculate
The local phase characteristic and the local amplitude characteristic of each pixel point in each distorted stereo image are obtained
Will be the local phase image and the local amplitude image of each distorted stereo image
The local phase image and the local amplitude image are associated as
And
then will be
The set of local phase images of all distorted stereo images in (1) is denoted as
And will be
The set of local amplitude image components of all distorted stereo images in (1) is expressed as
In this embodiment, in step ① _4,
and
the acquisition process comprises the following steps:
①_4a、using Log-Gabor filter pairs
Each pixel point in the image is filtered to obtain
The even symmetric frequency response and the odd symmetric frequency response of each pixel point in different scales and directions will be
The even symmetric frequency response of the pixel point with the middle coordinate position (x, y) in different scales and directions is recorded as e
α,θ(x, y) is
The odd symmetric frequency response of the pixel point with the middle coordinate position (x, y) in different scales and directions is recorded as o
α,θ(x, y), wherein x is more than or equal to 1 and less than or equal to W, y is more than or equal to 1 and less than or equal to H, α represents the scale factor of the Log-Gabor filter,
theta denotes a direction factor of the Log-Gabor filter,
① _4b, calculation
The phase consistency characteristics of each pixel point in different directions are
The phase consistency characteristics of the pixel points with the (x, y) middle coordinate position in different directions are recorded as PC
θ(x,y),
Wherein the content of the first and second substances,
① _4c, according to
The direction corresponding to the maximum phase consistency characteristic of each pixel point in the image is calculated
The local phase characteristic and the local amplitude characteristic of each pixel point in the image; for the
And (3) finding out the maximum phase consistency characteristic of the pixel point with the (x, y) middle coordinate position in the phase consistency characteristics in different directions, finding out the direction corresponding to the maximum phase consistency characteristic, and marking as theta
mAgain according to theta
mCalculating the local phase characteristic and the local amplitude characteristic of the pixel point, and correspondingly marking as
And
wherein the content of the first and second substances,
arctan () is an inverted cosine function,
to represent
The pixel point with the middle coordinate position (x, y) is in the direction theta corresponding to the different scales and the maximum phase consistency characteristics of the pixel point
mThe odd-symmetric frequency response of (a),
to represent
The pixel point with the middle coordinate position (x, y) is in the direction theta corresponding to the different scales and the maximum phase consistency characteristics of the pixel point
mThe even-symmetric frequency response of the frequency domain,
① _4d, according to
The local phase characteristics of all the pixel points in the image are obtained
Local phase image of
Also according to
Obtaining the local amplitude characteristics of all the pixel points in the image
Local amplitude image of
Obtaining according to steps ① _4a through ① _4d
And
in the same manner as the procedure of (1)
And
and
① _5, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the kth sub-block in all the local phase images in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein, the symbol
Is a sign of a lower rounding operation, k is more than or equal to 1 and less than or equal to M,
and
are each 64 x 1.
Also, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the kth sub-block in all the local phase images in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein the content of the first and second substances,
and
are each 64 x 1.
Also, will
Each local phase image of (1) and
each local amplitude image in (1) is divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in each local phase image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the kth sub-block in all the local phase images in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in each local amplitude image form the image characteristic vector of the sub-block in sequence, and the image characteristic vector is obtained
The image feature vector formed by the pixel values of all the pixel points in the k-th sub-block in all the local amplitude images in sequence is recorded as
Then will be
The set of image feature vectors of the sub-blocks in all the local phase images in (1) is denoted as
And will be
The set of image feature vectors of the sub-blocks in all the local amplitude images in (1) is expressed as
Wherein the content of the first and second substances,
and
are each 64 x 1.
① _6, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other; then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
1,k(ii) a Then will be
The set of image quality vectors of all the subblocks in all the distorted stereoscopic images in (1) is denoted by { y }
1,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
1,kHas a dimension of 6 x 1.
Also, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other; then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
2,k(ii) a Then will be
The set of image quality vectors of all the subblocks in all the distorted stereoscopic images in (1) is denoted by { y }
2,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
2,kHas a dimension of 6 x 1.
Also, will
Each distorted stereo image division in (1)
Sub-blocks with size of 8 × 8 and not overlapped with each other; then respectively obtaining the images by adopting 6 different full reference image quality evaluation methods
The objective evaluation prediction value of each subblock in each distorted stereo image is obtained; then will be
The 6 objective evaluation predicted values of each sub-block in each distorted stereo image form the image quality vector of the sub-block in sequence, and the image quality vector of the sub-block is obtained
The image quality vector formed by the 6 objective evaluation predicted values of the kth sub-block in all the distorted stereo images in sequence is recorded as y
3,k(ii) a Then will be
The set of image quality vectors of all the subblocks in all the distorted stereoscopic images in (1) is denoted by { y }
3,kL 1 is more than or equal to k and less than or equal to M }; wherein, y
3,kHas a dimension of 6 x 1.
In this embodiment, the 6 different full-reference image quality evaluation methods adopted are the known PSNR, MS-SSIM, FSIM, VIF, IW-SSIM, and UQI full-reference image quality evaluation methods, respectively.
① _7, pair of bags by K-SVD method
{y
1,k|1≤k≤M}、{y
2,kL 1 is less than or equal to k is less than or equal to M and y
3,kThe aggregate formed by |1 ≦ k ≦ M } is subjected to the joint dictionary training operation, and the structure is obtained
Respective image feature dictionary table and image quality dictionary table, corresponding
And
wherein the content of the first and second substances,
and
the dimensions of (a) are all 64 x K,
and
the dimensions of (a) are all 6 × K, K represents the number of set dictionaries, K is equal to or greater than 1, and in this embodiment, K is 256.
Also, the K-SVD method is adopted to pair
{y
1,k|1≤k≤M}、{y
2,kL 1 is less than or equal to k is less than or equal to M and y
3,kThe aggregate formed by |1 ≦ k ≦ M } is subjected to the joint dictionary training operation, and the structure is obtained
And
respective image feature dictionary table and image quality dictionary table, corresponding
And
wherein the content of the first and second substances,
and
the dimensions of (a) are all 64 x K,
and
the dimensions of (A) are all 6 XK.
In this embodiment, in step ① _7,
and
solving by adopting the existing K-SVD method
Obtained, where min () is a minimum function, the symbol "| | | | luminance
F"is a Frobenius norm-norm symbol for matrix calculation, the symbol" | | | | | | torry
1"is to find the 1-norm sign of the matrix, s is more than or equal to 1 and less than or equal to 3,
and
the dimensions of (a) are all 64 x M,
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
The mth first image feature vector in (1),
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
The mth first image feature vector in (1),
is composed of
The 1 st first image feature vector in (a),
is composed of
The kth first image feature vector in (1),
is composed of
M-th first image feature vector, Y
1=[y
1,1…y
1,k…y
1,M],Y
2=[y
2,1…y
2,k…y
2,M],Y
3=[y
3,1…y
3,k…y
3,M],Y
1、Y
2And Y
3All dimensions of (a) are 6 XM, y
1,1Is { y
1,k1 st image quality vector, y in |1 ≦ k ≦ M }
1,kIs { y
1,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
1,MIs { y
1,kMth image quality vector in |1 ≦ k ≦ M ≦ y
2,1Is { y
2,k1 st image quality vector, y in |1 ≦ k ≦ M }
2,kIs { y
2,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
2,MIs { y
2,kMth image quality vector in |1 ≦ k ≦ M ≦ y
3,1Is { y
3,k1 st image quality vector, y in |1 ≦ k ≦ M }
3,kIs { y
3,kK-th image quality vector in |1 ≦ k ≦ M ≦ y
3,MIs { y
3,kThe Mth image quality vector in |1 ≦ k ≦ M },
and
each of which represents a sparse matrix and each of which represents,
and
the dimensions of (A) are all K multiplied by M,
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
the dimension of (A) is K x 1, symbol [ "]]"is a vector symbol, γ is a weighting parameter, and γ is 0.5 and λ is a lagrangian parameter in this example, and λ is 0.15 in this example.
In step ① _7, the user selects,
and
the existing K-SVD method is adopted to solveSolution (II)
The process for preparing a novel compound of formula (I),
and
the dimensions of (a) are all 64 x M,
is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
is composed of
The 1 st second image feature vector in (b),
is composed of
The kth second image feature vector in (b),
is composed of
The mth second image feature vector in (1),
and
each of which represents a sparse matrix and each of which represents,
and
the dimensions of (A) are all K multiplied by M,
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
is composed of
The 1 st column vector of (1),
is composed of
The k-th column vector of (a),
is composed of
The M-th column vector of (1),
the dimensions of (A) are each K × 1.
The specific steps of the test phase process are as follows:
② _1, for any test stereo image S with width W' and height HtestWill StestIs recorded as LtestWill StestIs recorded as Rtest(ii) a Wherein W 'is the same as or different from W, and H' is the same as or different from H.
② _2, obtaining S in the same operation as step ① _4
test、L
testAnd R
testRespective local phase image and local amplitude image, and converting L into L
testThe local phase image and the local amplitude image are associated as
And
r is to be
testThe local phase image and the local amplitude image are associated as
And
② _3, will
And
are respectively divided into
Sub-blocks with size of 8 × 8 and not overlapped with each other; then will be
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
And will be
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Will be provided with
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Will be provided with
The pixel values of all pixel points in each sub-block in the image data are sequentially combined into an image characteristic vector of the sub-block, and the image characteristic vector is to be obtained
The image feature vector formed by the pixel values of all the pixel points in the t-th sub-block in sequence is recorded as
Then will be
The set of image feature vectors of all sub-blocks in (1) is denoted as
And will be
The set of image feature vectors of all sub-blocks in (1) is denoted as
Will be provided with
The set of image feature vectors of all sub-blocks in (1) is denoted as
Will be provided with
The set of image feature vectors of all sub-blocks in (1) is denoted as
Wherein the content of the first and second substances,
and
are each 64 x 1.
② _4 constructed according to the procedure in the training phase
Separately optimized reconstruction
And
a first sparse coefficient matrix for each respective image feature vector
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Also constructed from the process during the training phase
Separately optimized reconstruction
And
a second sparse coefficient matrix for each respective image feature vector
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Also constructed from the process during the training phase
Separately optimized reconstruction
And
a third sparse coefficient matrix for each respective image feature vector
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Also constructed from the process during the training phase
Separately optimized reconstruction
And
a first sparse coefficient matrix for each respective image feature vector
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a first sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Also constructed from the process during the training phase
Separately optimized reconstruction
And
second sparse coefficient of each image feature vector in eachMatrix of
Is expressed as a sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a second sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Also constructed from the process during the training phase
Separately optimized reconstruction
And
a third sparse coefficient matrix for each respective image feature vector
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
Obtained by
Is recorded as a third sparse coefficient matrix
Is solved by adopting a K-SVD method
And (4) obtaining the product.
Wherein the content of the first and second substances,
and
all dimensions of (a) are Kx 1, min () is a minimum function, and the symbol "| | | | non-conducting filament
F"is a Flobenius norm-norm symbol for solving the matrix, the symbol" | | | | | | luminance
1"is the 1-norm sign of the solved matrix, and λ is the Lagrangian parameter.
② _5 constructed according to the procedure in the training phase
Estimate separately
And
the first image quality vector of each sub-block in each will
The first image quality vector of the t sub-block of (1) is noted
Will be provided with
The first image quality vector of the t sub-block of (1) is noted
Also constructed from the process during the training phase
Estimate separately
And
the second image quality vector of each sub-block in each will
The second image quality vector of the t sub-block of (1)
Will be provided with
Second image quality of the t sub-block in (1)The vector is noted as
Also constructed from the process during the training phase
Estimate separately
And
the third image quality vector of each sub-block in each
The third image quality vector of the t sub-block of (1)
Will be provided with
The third image quality vector of the t sub-block of (1)
Also constructed from the process during the training phase
Estimate separately
And
the first image quality vector of each sub-block in each will
The first image quality vector of the t sub-block of (1) is noted
Will be provided with
The first image quality vector of the t sub-block of (1) is noted
Also constructed from the process during the training phase
Estimate separately
And
the second image quality vector of each sub-block in each will
The second image quality vector of the t sub-block of (1)
Will be provided with
The second image quality vector of the t sub-block of (1)
Also constructed from the process during the training phase
Estimate separately
And
the third image quality vector of each sub-block in each
The third image quality vector of the t sub-block of (1)
Will be provided with
The third image quality vector of the t sub-block of (1)
Wherein the content of the first and second substances,
and
the dimensions of (a) are all 6 × 1.
② _6, calculation
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein exp () represents an exponential function with a natural base e as a base, and the symbol "| | | | purple
2"2-norm sign of matrix is obtained, η is control parameter, η is 1000 in this embodiment,
is composed of
The input vector of (1).
Also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
In (1)The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1).
Also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1).
Also, calculate
And the multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality of each sub-block in the image processing system
The multi-distortion fusion sparse coefficient matrix and the multi-distortion fusion image quality correspondence of the t sub-block in the sequence are recorded as
And
wherein the content of the first and second substances,
is composed of
The input vector of (1).
② _7, calculation
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
L,P,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
R,P,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
L,A,
Also, calculate
And the global sparse coefficient matrix and the global image quality are correspondingly recorded as
And Q
R,A,
② _8, according to
And
and Q
L,PAnd Q
R,PCalculating S
testThe predicted value of the objective quality evaluation of the local phase image is recorded as Q
P,Q
P=ω
L,P×Q
L,P+ω
R,P×Q
R,P(ii) a Wherein, ω is
L,PIs Q
L,PThe weight of (a) is calculated,
ω
R,Pis Q
R,PThe weight of (a) is calculated,
symbol'<>"is the inner product symbol, C is the control parameter, and in this example, C is 0.02.
Also according to
And
and Q
L,AAnd Q
R,ACalculating S
testThe predicted value of the objective evaluation of the quality of the local amplitude image is marked as Q
A,Q
A=ω
L,A×Q
L,A+ω
R,A×Q
R,A(ii) a Wherein, ω is
L,AIs Q
L,AThe weight of (a) is calculated,
ω
R,Ais Q
R,AThe weight of (a) is calculated,
② _9, according to QPAnd QACalculating StestThe predicted value of the objective evaluation of image quality is expressed as Q, Q ═ ω (ω)P×(QP)n+(1-ωP)×(QA)n)1/n(ii) a Wherein, ω isPAnd n are weighted parameters, in this example, take ωP=0.35、n=1。
In this embodiment, an asymmetric multi-distortion stereo image database established by Ningbo university is used to analyze the correlation between the image quality objective evaluation prediction value of the distorted stereo image obtained in this embodiment and the average subjective score difference value, where the asymmetric multi-distortion stereo image database established by Ningbo university includes 3000 asymmetric multi-distortion stereo images, and the average subjective score difference value of each distorted stereo image in the asymmetric multi-distortion stereo image database is obtained by using a subjective quality evaluation method.
In this embodiment, 3 common objective parameters of the evaluation method for evaluating image quality are used as evaluation indexes, that is, a Pearson correlation coefficient (PLCC), a Spearman correlation coefficient (SROCC), a mean square error (RMSE), and the PLCC and the RMSE reflect the accuracy of the objective evaluation model for distorted stereoscopic images, and the SROCC reflects monotonicity thereof under a nonlinear regression condition. The Pearson correlation coefficient, the Spearman correlation coefficient and the mean square error between the image quality objective evaluation predicted value of the distorted three-dimensional image and the average subjective score difference value, which are obtained by respectively adopting the method of the invention and the known PSNR and SSIM full reference quality evaluation methods, are compared, and the comparison result is shown in table 1, and the table 1 shows that the correlation between the image quality objective evaluation predicted value of the distorted three-dimensional image obtained by adopting the method of the invention and the average subjective score difference value is very high, so that the objective evaluation result of the method of the invention is fully consistent with the result of human eye subjective perception, and the feasibility and the effectiveness of the method of the invention can be sufficiently demonstrated.
TABLE 1 Pearson correlation coefficient comparison, Spearman correlation coefficient comparison, and mean square error comparison between objective evaluation predicted value of image quality and mean subjective score difference of distorted stereoscopic images obtained by the method of the present invention and the known full-reference quality evaluation method
|
Pearson correlation coefficient
|
Spearman correlation coefficient
|
Mean square error
|
PSNR method
|
0.7003
|
0.7139
|
8.4165
|
SSIM method
|
0.7144
|
0.7339
|
8.1642
|
The method of the invention
|
0.7853
|
0.7652
|
7.4416 |