One kind is without with reference to stereo image quality evaluation method
Technical field
The present invention relates to a kind of image quality evaluating method, more particularly, to one kind without with reference to stereo image quality evaluation side
Method.
Background technology
With developing rapidly for the technology such as Image Coding and display, image quality evaluation research has become wherein to be weighed very much
The link wanted.The target of method for objectively evaluating image quality research is consistent as far as possible with subjective evaluation result, so as to put
De- time-consuming and uninteresting picture quality subjective evaluation method, it can utilize computer automatically evaluation image quality.According to right
The reference of original image and degree of dependence, method for objectively evaluating image quality can be divided into three major types:It is complete to refer to (Full
Reference, FR) image quality evaluating method, partly refer to (Reduced Reference, RR) image quality evaluating method
With without refer to (No Reference, NR) image quality evaluating method.
Non-reference picture quality appraisement method is due to without any reference image information, with higher flexibility, therefore
Receive more and more extensive concern.At present, existing non-reference picture quality appraisement method is predicted by machine learning
Evaluation model, but its computation complexity is higher, and training pattern needs to predict the subjective assessment value of each evaluation image, not
Suitable for actual application scenario, have some limitations.Rarefaction representation is the effective way of evaluation image quality, and key exists
In how effectively constructing the substantive characteristics that dictionary carrys out phenogram picture, and for stereo-picture, left view dot image and right viewpoint figure
The symmetrical and asymmetric of picture can cause different binocular vision characteristics, therefore, how cause construction dictionary have distinguishability,
How contact is set up between characteristics of image and the dictionary of picture quality, be all that reference-free quality evaluation is being carried out to stereo-picture
The technical problem for needing emphasis to solve in research.
The content of the invention
The technical problems to be solved by the invention are to provide one kind without stereo image quality evaluation method is referred to, and it can have
Effect ground improves the correlation between objective evaluation result and subjective perception, and need not predict the subjective assessment value of each evaluation image.
The present invention solve the technical scheme that is used of above-mentioned technical problem for:One kind is without with reference to stereo image quality evaluation side
Method, it is characterised in that including two processes of training stage and test phase;
Described training stage process is comprised the following steps that:
1. the original undistorted stereo-picture that N breadth degree is W and height is H _ 1, is chosen, it is original undistorted by the u
Stereo-picture is designated asWillLeft view dot image corresponding with right visual point image be designated asWithThen obtain every it is original
Undistorted stereo-picture one-eyed figure, willOne-eyed seal beThen the undistorted stereo-picture original to every
The one-eyed figure of left view dot image and right visual point image and every original undistorted stereo-picture carries out L different strength of distortion respectively
Distortion, the distortion of the left view dot image of all original undistorted stereo-pictures and each self-corresponding L strength of distortion is left
Visual point image constitutes first group of training image collection, is designated asAnd will be all original without mistake
The right visual point image of distortion of the right visual point image of true stereo-picture and each self-corresponding L strength of distortion constitutes second group of training figure
Image set, is designated asBy the one-eyed figure of all original undistorted stereo-pictures and each it is right
The one-eyed figure of distortion for the L strength of distortion answered constitutes the 3rd group of training image collection, is designated as
Wherein, N>1,1≤u≤N, L>1,1≤v≤L,RepresentThe distortion left view dot image of corresponding v-th of strength of distortion,RepresentThe right visual point image of distortion of corresponding v-th of strength of distortion,RepresentCorresponding v-th of strength of distortion
The one-eyed figure of distortion;
1. _ 2, willIn every width distortion left view dot image be divided intoIt is individual mutual
Nonoverlapping size is 16 × 16 sub-block;Then obtained using two kinds of different natural scene statistical methodsIn every width distortion left view dot image in each sub-block characteristics of image vector, willIn all distortion left view dot images in the characteristics of image vector of k-th of sub-block be designated asAgain willIn all distortion left view dot images in the respective image of all sub-blocks
The set that characteristic vector is constituted is designated as
And willIn every right visual point image of width distortion be divided intoIt is individual mutually not
Overlapping size is 16 × 16 sub-block;Then obtained using two kinds of different natural scene statistical methodsIn every right visual point image of width distortion in each sub-block characteristics of image vector, willIn the right visual point image of all distortions in k-th of sub-block characteristics of image vector note
ForAgain willIn the right visual point image of all distortions in all sub-blocks it is respective
The set that characteristics of image vector is constituted is designated as
WillIn every one-eyed figure of width distortion be divided intoThe size of individual non-overlapping copies
Size is 16 × 16 sub-block;Then obtained using two kinds of different natural scene statistical methods
In every one-eyed figure of width distortion in each sub-block characteristics of image vector, willIn institute
The characteristics of image vector for having k-th of sub-block in the one-eyed figure of distortion is designated asAgain willIn
The one-eyed figure of all distortions in the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated as
Wherein,1≤k≤M,WithDimension be 72 × 1;
1. _ 3, obtained respectively using six kinds of different full reference image quality appraisement methods
In every width distortion left view dot image in each sub-block objective evaluation predicted value;Then will
In every width distortion left view dot image in 6 objective evaluation predicted values of each sub-block sequentially constitute the picture quality of the sub-block
Vector, willIn all distortion left view dot images in 6 of k-th of sub-block objective comment
The picture quality vector that valency predicted value is sequentially constituted is designated asAgain willIn lost
The set that the respective picture quality vector of all sub-blocks in true left view dot image is constituted is designated as
And obtained respectively using six kinds of different full reference image quality appraisement methods
In every right visual point image of width distortion in each sub-block objective evaluation predicted value;Then will
In every right visual point image of width distortion in 6 objective evaluation predicted values of each sub-block sequentially constitute the picture quality of the sub-block
Vector, willIn the right visual point image of all distortions in 6 of k-th of sub-block objective comment
The picture quality vector that valency predicted value is sequentially constituted is designated asAgain willIn lost
The set that the respective picture quality vector of all sub-blocks in very right visual point image is constituted is designated as
Obtained respectively using six kinds of different full reference image quality appraisement methodsIn
Every one-eyed figure of width distortion in each sub-block objective evaluation predicted value;Then willIn
Every one-eyed figure of width distortion in 6 objective evaluation predicted values of each sub-block sequentially constitute the picture quality vector of the sub-block,
WillIn the one-eyed figure of all distortions in k-th of sub-block 6 objective evaluation predicted values
The picture quality vector sequentially constituted is designated asAgain willIn the one-eyed figure of all distortions
In the set that constitutes of the respective picture quality vector of all sub-blocks be designated as
Wherein,WithDimension be 6 × 1;
1. _ 4, using K-SVD methods to by WithThe set of composition carries out joint dictionary training operation, structure
Make and obtainWith
Respective characteristics of image dictionary table and picture quality dictionary table, willCharacteristics of image dictionary
Table is corresponding with picture quality dictionary table to be designated as DLAnd WL, willCharacteristics of image dictionary table and
Picture quality dictionary table correspondence is designated as DRAnd WR, willCharacteristics of image dictionary table and image
Quality dictionary table correspondence is designated as DCAnd WC;Wherein, DL、DRAnd DCDimension be 72 × K, WL、WRAnd WCDimension be 6 × K, K
Represent the number of the dictionary of setting, K >=1;
Described test phase process is comprised the following steps that:
2. the test distortion stereo-picture S for _ 1, being W' for any one breadth degree and being highly H'test, by StestLeft view
Dot image is corresponding with right visual point image to be designated as LtestAnd Rtest;Then S is obtainedtestOne-eyed figure, be designated as Ctest;Then by Ltest、
RtestAnd CtestIt is divided intoThe size of individual non-overlapping copies is 16 × 16 sub-block;
2. _ 2, according to step 1. _ 2 in obtainWithProcess, with identical operation obtain Ltest、
RtestAnd CtestThe characteristics of image vector of each sub-block in each, by LtestIn the characteristics of image vector of t-th of sub-block be designated asBy RtestIn the characteristics of image vector of t-th of sub-block be designated asBy CtestIn t-th of sub-block characteristics of image vector
It is designated asThen by LtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asAnd by RtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asBy CtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asWherein, WithDimension be 72 × 1;
2. _ 3, according to the D obtained in training stage procedure constructionL、DRAnd DC, obtained by combined optimizationWithEach characteristics of image arrow in each
The sparse coefficient matrix of amount, willIn the sparse coefficient matrix of t-th of characteristics of image vector be designated asWillIn the sparse coefficient matrix of t-th of characteristics of image vector be designated asWillIn the sparse coefficient matrix of t-th of characteristics of image vector be designated as WithIt is to pass through
SolveObtain;Again willIn all characteristics of image arrow
The set that the sparse coefficient matrix of amount is constituted is designated asWillIn all figures
As the set that the sparse coefficient matrix of characteristic vector is constituted is designated asWillIn
All characteristics of image vectors sparse coefficient matrix constitute set be designated asWherein,
WithDimension K × 1, min () is to take minimum value function, symbol " | | | |F" be ask for the not Luo Beini Wu Si of matrix-
Norm sign;
2. _ 4, according to the W obtained in training stage procedure constructionL, estimate LtestIn each sub-block picture quality arrow
Amount, by LtestIn the picture quality vector of t-th of sub-block be designated as And according in training stage process structure
Make obtained WR, estimate RtestIn each sub-block picture quality vector, by RtestIn t-th of sub-block picture quality arrow
Amount is designated as According to the W obtained in training stage procedure constructionC, estimate CtestIn each sub-block image
Quality vector, by CtestIn the picture quality vector of t-th of sub-block be designated as Wherein,With
Dimension be 6 × 1;
2. L _ 5, is calculatedtestPicture quality objective evaluation predicted value, be designated as qL,And calculate
RtestPicture quality objective evaluation predicted value, be designated as qR,Calculate CtestPicture quality objective comment
Valency predicted value, is designated as qC,Wherein,Represent LtestIn t-th of sub-block in all pixels points
The standard deviation of pixel value,Represent RtestIn t-th of sub-block in all pixels point pixel value standard deviation,Represent
CtestIn t-th of sub-block in all pixels point pixel value standard deviation, symbol " | | | |1" it is the 1- norms for asking for matrix
Symbol;
2. _ 6, basisWithIn
The sparse coefficient matrix of all characteristics of image vectors, obtains Ltest、RtestAnd CtestRespective weight coefficient, correspondence is designated as WithThen S is calculatedtestPicture quality objective evaluation predicted value, be designated as Q,
Described step 1. _ 2 inAcquisition process be:Obtained using BRISQUE methodsIn all distortion left view dot images in k-th of sub-block the first characteristics of image arrow
Amount, and obtained using LBP methodsIn all distortion left view dot images in k-th son
Second characteristics of image vector of block, wherein, the dimension of first characteristics of image vector the second characteristics of image vector is 36 × 1;
Then willIn all distortion left view dot images in k-th of sub-block the first image it is special
Vector the second characteristics of image vector is levied sequentially to be combined into
Described step 1. _ 2 inAcquisition process be:Obtained using BRISQUE methodsIn the right visual point image of all distortions in k-th of sub-block the first characteristics of image arrow
Amount, and obtained using LBP methodsIn the right visual point image of all distortions in k-th
Second characteristics of image vector of sub-block, wherein, the dimension of first characteristics of image vector the second characteristics of image vector is 36 ×
1;Then willIn the right visual point image of all distortions in k-th of sub-block the first figure
As characteristic vector and the second characteristics of image vector are sequentially combined into
Described step 1. _ 2 inAcquisition process be:Obtained using BRISQUE methodsIn the one-eyed figure of all distortions in k-th of sub-block the first characteristics of image vector, and
Obtained using LBP methodsIn the one-eyed figure of all distortions in k-th of sub-block second
Characteristics of image vector, wherein, the dimension of first characteristics of image vector the second characteristics of image vector is 36 × 1;Then willIn the one-eyed figure of all distortions in k-th of sub-block the first characteristics of image vector
Second characteristics of image vector is sequentially combined into
Described step 1. _ 3 in six kinds of different full reference image quality appraisement methods using be respectively PSNR,
The full reference image quality appraisement method of SSIM, PC, GM, VIF and UQI.
Described step 1. _ 4 in DL、DR、DC、WL、WRAnd WCIt is to be solved using K-SVD methodsObtain, wherein, min
() is to take minimum value function, symbol " | | | |F" it is to ask for not Luo Beini Wu Si-norm sign of matrix, symbol " | | | |1" be
The 1- norm signs of matrix are asked for, XL、XRAnd XCDimension be 72 × M,Represent
In all distortion left view dot images in the 1st sub-block characteristics of image vector,Represent In all distortion left view dot images in m-th sub-block characteristics of image vector,RepresentIn the right visual point image of all distortions in the 1st sub-block characteristics of image vector,
RepresentIn the right visual point image of all distortions in m-th sub-block characteristics of image arrow
Amount,RepresentIn the one-eyed figure of all distortions in the 1st sub-block characteristics of image
Vector,RepresentIn the one-eyed figure of all distortions in m-th sub-block image it is special
Levy vector,
YL、YRAnd YCDimension be 6 × M,RepresentIn all distortion left view dot images
In the 1st sub-block the picture quality vector that sequentially constitutes of 6 objective evaluation predicted values,RepresentIn all distortion left view dot images in m-th sub-block 6 objective evaluations prediction
The picture quality vector that value is sequentially constituted,RepresentIn the right viewpoint of all distortions
The picture quality vector that 6 objective evaluation predicted values of the 1st sub-block in image are sequentially constituted,RepresentIn the right visual point image of all distortions in m-th sub-block 6 objective evaluations prediction
The picture quality vector that value is sequentially constituted,RepresentIn the one-eyed figure of all distortions
In the 1st sub-block the picture quality vector that sequentially constitutes of 6 objective evaluation predicted values,RepresentIn the one-eyed figure of all distortions in m-th sub-block 6 objective evaluations it is pre-
The picture quality vector that measured value is sequentially constituted, AL、ARAnd ACIt is sparse matrix, AL、ARAnd ACDimension be K × M,
For ALIn the 1st column vector,For ALIn k-th of column vector,For ALIn m-th column vector,For ARIn
The 1st column vector,For ARIn k-th of column vector,For ARIn m-th column vector,For ACIn the 1st
Individual column vector,For ACIn k-th of column vector,For ACIn m-th column vector, WithDimension be K × 1, symbol " [] " is vector representation symbol, and α joins for weighting
Number, λ is LaGrange parameter.
Described step 2. _ 6 inWithAcquisition process be:
A1, calculatingIn all characteristics of image vectors sparse coefficient matrix histogram, note
For
And calculateIn all characteristics of image vectors sparse coefficient matrix histogram, note
For
CalculateIn all characteristics of image vectors sparse coefficient matrix histogram, be designated as
Wherein, 1≤g≤K, P are representedIn Nogata node of graph sum, also represent
In Nogata node of graph sum, also representIn Nogata node of graph sum, it with 2 is bottom that log (), which is represented,
Logarithmic function, 1≤p≤P,RepresentIn belong to the probability density function of g-th of dictionary component,
RepresentIn belong to the probability density function of g-th of dictionary component,Represent
In belong to the probability density function of g-th of dictionary component,RepresentIn belong to p-th of histogram
The probability density function of node,RepresentIn belong to the probability density letter of p-th of Nogata node of graph
Number,RepresentIn belong to the probability density function of p-th of Nogata node of graph,RepresentIn
G-th of dictionary component it is quantified after coefficient value,RepresentIn g-th of dictionary component it is quantified after coefficient
Value,RepresentIn g-th of dictionary component it is quantified after coefficient value,
A2, calculatingWithIn all characteristics of image vectors it is sparse
The joint histogram of coefficient matrix, is designated as Wherein, P is also representedIn it is straight
The sum of square node of graph, 1≤p1≤ P, 1≤p2≤ P,RepresentWithIn
Belong to the joint probability density function of g-th of dictionary component,RepresentIn belong to pth1It is individual straight
Square node of graph and pth2The joint probability density function of individual Nogata node of graph,
A3, calculating Ltest、RtestAnd CtestRespective independent entropy, correspondence is designated as H (Ltest)、H(Rtest) and H (Ctest),Then count
Calculate LtestAnd RtestCombination entropy, be designated as H (Ltest,Rtest),
A4, calculating Ltest、RtestAnd CtestRespective weight coefficient, is corresponded toWith
Compared with prior art, the advantage of the invention is that:
1) the inventive method respectively obtains all original undistorted stereograms in the training stage by combining dictionary training
First group of training image collection of the distortion left view point image construction of the left view dot image of picture and each self-corresponding different strength of distortion,
The right visual point image of distortion of the right visual point image of all original undistorted stereo-pictures and each self-corresponding different strength of distortion
Second group of training image collection, the one-eyed figure of all original undistorted stereo-pictures and each self-corresponding different distortions constituted are strong
The 3rd group of respective characteristics of image dictionary table of training image collection and picture quality dictionary table that the one-eyed figure of distortion of degree is constituted;Surveying
The examination stage need not calculate characteristics of image dictionary table and picture quality dictionary table again, and this avoid complicated machine learning training
Process, and the subjective assessment value of each test distortion stereo-picture need not be predicted so that the inventive method is applied to actual answer
Use occasion.
2) the inventive method is in test phase, according to first group of training image collection, second obtained in training stage construction
Group training image collection and the 3rd group of respective characteristics of image dictionary table of training image collection, obtain test distortion by combined optimization and stand
Left view dot image, right visual point image and the one-eyed figure of body image each in each sub-block characteristics of image vector sparse coefficient
Matrix;And the sparse coefficient square of the characteristics of image vector of each sub-block in the left view dot image for passing through test distortion stereo-picture
It is every in the picture quality dictionary table of battle array and first group of training image collection, the left view dot image to obtain test distortion stereo-picture
The picture quality vector of individual sub-block, so that the picture quality objective evaluation for obtaining the left view dot image of test distortion stereo-picture is pre-
Measured value;Pass through the sparse coefficient matrix of the characteristics of image vector of each sub-block in the right visual point image of test distortion stereo-picture
It is each in the right visual point image to obtain test distortion stereo-picture with the picture quality dictionary table of second group of training image collection
The picture quality vector of sub-block, so as to obtain the picture quality objective evaluation prediction of the right visual point image of test distortion stereo-picture
Value;Pass through the sparse coefficient matrix and the 3rd of the characteristics of image vector of each sub-block in the one-eyed figure of test distortion stereo-picture
The picture quality dictionary table of group training image collection, the image of each sub-block in one-eyed figure to obtain test distortion stereo-picture
Quality vector, so as to obtain the picture quality objective evaluation predicted value of the one-eyed figure of test distortion stereo-picture;Consider further that to a left side
Visual point image, right visual point image and the respective weight coefficient of one-eyed figure, to obtain the picture quality visitor of test distortion stereo-picture
Evaluation and foreca value is seen, obtained objective evaluation result maintains preferable uniformity with subjective assessment value.
Brief description of the drawings
Fig. 1 realizes block diagram for the totality of the inventive method.
Embodiment
The present invention is described in further detail below in conjunction with accompanying drawing embodiment.
It is proposed by the present invention a kind of without stereo image quality evaluation method is referred to, its totality realizes block diagram as shown in figure 1, its
Including two processes of training stage and test phase.
Described training stage process is comprised the following steps that:
1. the original undistorted stereo-picture that N breadth degree is W and height is H _ 1, is chosen, it is original undistorted by the u
Stereo-picture is designated asWillLeft view dot image corresponding with right visual point image be designated asWithThen using existing
The one-eyed figure of the original undistorted stereo-picture of technical limit spacing every, willOne-eyed seal beThen it is original to every
The one-eyed figure difference of the left view dot image and right visual point image of undistorted stereo-picture and every original undistorted stereo-picture
The distortion of L different strength of distortion is carried out, by the left view dot image and each self-corresponding L of all original undistorted stereo-pictures
First group of training image collection of distortion left view point image construction of individual strength of distortion, is designated asAnd
By the right visual point image of all original undistorted stereo-pictures and the right visual point image of distortion of each self-corresponding L strength of distortion
Second group of training image collection is constituted, is designated asBy all original undistorted stereo-pictures
One-eyed figure and the one-eyed figure of distortion of each self-corresponding L strength of distortion constitute the 3rd group of training image collection, be designated asWherein, N>1, N=200,1≤u≤N, L are taken in the present embodiment>1, in this implementation
L=5,1≤v≤L are taken in example,RepresentThe distortion left view dot image of corresponding v-th of strength of distortion,RepresentIt is right
The right visual point image of distortion for v-th of the strength of distortion answered,RepresentThe one-eyed figure of distortion of corresponding v-th of strength of distortion,
Symbol " { } " is set expression symbol.
1. _ 2, willIn every width distortion left view dot image be divided intoIt is individual mutual
Nonoverlapping size is 16 × 16 sub-block;Then using existing two kinds different natural scene statistics (natural
Scene statistics) method acquisitionIn every width distortion left view dot image in it is every
The characteristics of image vector of individual sub-block, willIn all distortion left view dot images in kth
The characteristics of image vector of individual sub-block is designated asAgain willIn all distortion left view point diagrams
The set that the respective characteristics of image vector of all sub-blocks as in is constituted is designated as
And willIn every right visual point image of width distortion be divided intoIt is individual mutually not
Overlapping size is 16 × 16 sub-block;Then using existing two kinds different natural scene statistics (natural
Scene statistics) method acquisitionIn every right visual point image of width distortion in
The characteristics of image vector of each sub-block, willIn the right visual point image of all distortions in
The characteristics of image vector of k-th of sub-block is designated asAgain willIn all distortion right sides regard
The set that the respective characteristics of image vector of all sub-blocks in dot image is constituted is designated as
WillIn every one-eyed figure of width distortion be divided intoIndividual non-overlapping copies
Size is 16 × 16 sub-block;Then using existing two kinds different natural scene statistics (natural scene
Statistics) method is obtainedIn every one-eyed figure of width distortion in each sub-block
Characteristics of image vector, willIn the one-eyed figure of all distortions in k-th of sub-block image
Characteristic vector is designated asAgain willIn the one-eyed figure of all distortions in all sub-blocks
The set that respective characteristics of image vector is constituted is designated as
Wherein,1≤k≤M,WithDimension be 72 × 1.
In the present embodiment, step 1. _ 2 in two kinds of different natural scene statistical methods be respectively BRISQUE
(Blind/Referenceless Image Spatial QUality Evaluator) method and LBP (local binary
Pattern) method.Step 1. _ 2 inAcquisition process be:Obtained using BRISQUE methodsIn all distortion left view dot images in k-th of sub-block the first characteristics of image arrow
Amount, and obtained using LBP methodsIn all distortion left view dot images in k-th son
Second characteristics of image vector of block, wherein, the dimension of first characteristics of image vector the second characteristics of image vector is 36 × 1;
Then willIn all distortion left view dot images in k-th of sub-block the first image it is special
Vector the second characteristics of image vector is levied sequentially to be combined intoStep 1. _ 2 inAcquisition process be:Using
BRISQUE methods are obtainedIn all distortions right visual point images in k-th of sub-block
First characteristics of image vector, and obtained using LBP methodsIn the right viewpoint of all distortions
Second characteristics of image vector of k-th of sub-block in image, wherein, first characteristics of image vector the second characteristics of image vector
Dimension is 36 × 1;Then willIn the right visual point image of all distortions in k-th
First characteristics of image vector the second characteristics of image vector of sub-block is sequentially combined intoStep 1. _ 2 inAcquisition
Process is:Obtained using BRISQUE methodsIn the one-eyed figure of all distortions in kth
First characteristics of image vector of individual sub-block, and obtained using LBP methodsIn lost
Second characteristics of image vector of k-th of sub-block in very one-eyed figure, wherein, first the second characteristics of image of characteristics of image vector arrow
The dimension of amount is 36 × 1;Then willIn the one-eyed figure of all distortions in k-th
First characteristics of image vector the second characteristics of image vector of sub-block is sequentially combined into
1. _ 3, obtained respectively using existing six kinds different full reference image quality appraisement methodsIn every width distortion left view dot image in each sub-block objective evaluation predicted value;So
Afterwards willIn every width distortion left view dot image in each sub-block 6 objective evaluations it is pre-
Measured value sequentially constitutes the picture quality vector of the sub-block, willIn all distortion left view points
The picture quality vector that 6 objective evaluation predicted values of k-th of sub-block in image are sequentially constituted is designated asAgain willIn all distortion left view dot images in the respective picture quality vector of all sub-blocks
The set of composition is designated as
And obtained respectively using existing six kinds different full reference image quality appraisement methods In every right visual point image of width distortion in each sub-block objective evaluation predicted value;Then will In every right visual point image of width distortion in 6 objective evaluation predicted values of each sub-block sequentially constitute
The picture quality vector of the sub-block, willIn the right visual point image of all distortions in kth
The picture quality vector that 6 objective evaluation predicted values of individual sub-block are sequentially constituted is designated asAgain will In the right visual point image of all distortions in the set that constitutes of the respective picture quality vector of all sub-blocks be designated as
Obtained respectively using existing six kinds different full reference image quality appraisement methods In every one-eyed figure of width distortion in each sub-block objective evaluation predicted value;Then will In every one-eyed figure of width distortion in 6 objective evaluation predicted values of each sub-block sequentially constitute the image of the sub-block
Quality vector, willIn the one-eyed figure of all distortions in 6 of k-th of sub-block it is objective
The picture quality vector that evaluation and foreca value is sequentially constituted is designated asAgain willIn institute
The set for having the respective picture quality vector of all sub-blocks in the one-eyed figure of distortion to constitute is designated as
Wherein,WithDimension be 6 × 1.
In the present embodiment, step 1. _ 3 in six kinds of different full reference image quality appraisement methods using be respectively public affairs
The PSNR (peak signal to noise ratio) that knows, SSIM (structural similarity index), PC
(phase congruency from feature similarity index)、GM(gradient magnitude from
Feature similarity index), VIF (visual information fidelity) and UQI (universal
Quality index) full reference image quality appraisement method.Step 1. _ 3 inAcquisition process be:Using six kinds not
Same full reference image quality appraisement method is obtained respectivelyIn all distortion left view points
The objective evaluation predicted value of k-th of sub-block in image, is obtainedIn all distortions it is left
6 objective evaluation predicted values of k-th of sub-block in visual point image;Then willIn institute
6 objective evaluation predicted values for having k-th of sub-block in distortion left view point image are sequentially combined into the picture quality vector of the sub-blockStep 1. _ 3 inAcquisition process be:Obtained respectively using six kinds of different full reference image quality appraisement methods
TakeIn the right visual point image of all distortions in k-th of sub-block objective evaluation prediction
Value, is obtainedIn the right visual point image of all distortions in 6 of k-th of sub-block it is objective
Evaluation and foreca value;Then willIn all distortions right visual point images in k-th of sub-block
6 objective evaluation predicted values are sequentially combined into the picture quality vector of the sub-blockStep 1. _ 3 inAcquisition
Cheng Wei:Obtained respectively using six kinds of different full reference image quality appraisement methodsIn
The objective evaluation predicted value of k-th of sub-block in all one-eyed figures of distortion, is obtainedIn
6 objective evaluation predicted values of k-th of sub-block in all one-eyed figures of distortion;Then will
In the one-eyed figure of all distortions in 6 objective evaluation predicted values of k-th of sub-block be sequentially combined into the picture quality of the sub-block
Vector
1. _ 4, using existing K-SVD methods to by WithThe set of composition carries out joint dictionary training operation, structure
Make and obtain With
Respective characteristics of image dictionary table and picture quality dictionary table, willCharacteristics of image dictionary
Table is corresponding with picture quality dictionary table to be designated as DLAnd WL, willCharacteristics of image dictionary table and
Picture quality dictionary table correspondence is designated as DRAnd WR, willCharacteristics of image dictionary table and image
Quality dictionary table correspondence is designated as DCAnd WC;Wherein, DL、DRAnd DCDimension be 72 × K, WL、WRAnd WCDimension be 6 × K, K
The number of the dictionary of setting is represented, K >=1 takes K=256 in the present embodiment.
In this particular embodiment, step 1. _ 4 in DL、DR、DC、WL、WRAnd WCIt is to be asked using existing K-SVD methods
SolutionObtain, wherein, min ()
To take minimum value function, symbol " | | | |F" it is this black (the Frobeniu)-norm sign of not Luo Beini for asking for matrix, symbol " |
| ||1" it is the 1- norm signs for asking for matrix, XL、XRAnd XCDimension be 72 × M,RepresentIn
All distortion left view dot images in the 1st sub-block characteristics of image vector,Represent
In all distortion left view dot images in m-th sub-block characteristics of image vector,Represent
In the right visual point image of all distortions in the 1st sub-block characteristics of image vector,Represent
In the right visual point image of all distortions in m-th sub-block characteristics of image vector,Represent
In the one-eyed figure of all distortions in the 1st sub-block characteristics of image vector,Represent
In the one-eyed figure of all distortions in m-th sub-block characteristics of image vector, YL、YRAnd YCDimension be 6 × M,
RepresentIn all distortion left view dot images in the 1st sub-block 6 objective evaluations
The picture quality vector that predicted value is sequentially constituted,RepresentIn all distortion left views
The picture quality vector that 6 objective evaluation predicted values of the m-th sub-block in dot image are sequentially constituted,RepresentIn the right visual point image of all distortions in the 1st sub-block 6 objective evaluations prediction
The picture quality vector that value is sequentially constituted,RepresentIn the right viewpoint of all distortions
The picture quality vector that 6 objective evaluation predicted values of the m-th sub-block in image are sequentially constituted,RepresentIn the one-eyed figure of all distortions in 6 objective evaluation predicted values of the 1st sub-block press
The picture quality vector of sequence composition,RepresentIn the one-eyed figure of all distortions in
The picture quality vector that 6 objective evaluation predicted values of m-th sub-block are sequentially constituted, AL、ARAnd ACIt is sparse matrix, AL、ARWith
ACDimension be K × M,For ALIn the 1st column vector,For ALIn k-th of column vector,For ALIn
M column vector,For ARIn the 1st column vector,For ARIn k-th of column vector,For ARIn m-th arrange to
Amount,For ACIn the 1st column vector,For ACIn k-th of column vector,For ACIn m-th column vector, WithDimension be K × 1, symbol " [] "
For vector representation symbol, α is weighting parameters, and it is LaGrange parameter that α=0.5, λ is taken in the present embodiment, in the present embodiment
Take λ=0.15.
Described test phase process is comprised the following steps that:
2. the test distortion stereo-picture S for _ 1, being W' for any one breadth degree and being highly H'test, by StestLeft view
Dot image is corresponding with right visual point image to be designated as LtestAnd Rtest;Then existing technical limit spacing S is usedtestOne-eyed figure, be designated as
Ctest;Then by Ltest、RtestAnd CtestIt is divided intoThe size of individual non-overlapping copies is 16 × 16 sub-block;
Wherein, W' and W, H' and H can be with identical, can also be different.
2. _ 2, according to step 1. _ 2 in obtainWithProcess, with identical operation obtain Ltest、
RtestAnd CtestThe characteristics of image vector of each sub-block in each, by LtestIn the characteristics of image vector of t-th of sub-block be designated asBy RtestIn the characteristics of image vector of t-th of sub-block be designated asBy CtestIn t-th of sub-block characteristics of image vector
It is designated asThen by LtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asAnd by RtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asBy CtestIn the set that constitutes of the respective characteristics of image vector of all sub-blocks be designated asWherein, WithDimension be 72 × 1.
2. _ 3, according to the D obtained in training stage procedure constructionL、DRAnd DC, obtained by combined optimizationWithEach characteristics of image arrow in each
The sparse coefficient matrix of amount, willIn the sparse coefficient matrix of t-th of characteristics of image vector be designated asWillIn the sparse coefficient matrix of t-th of characteristics of image vector be designated asWillIn the sparse coefficient matrix of t-th of characteristics of image vector be designated asWithIt is to pass through
SolveObtain;Again willIn all characteristics of image arrow
The set that the sparse coefficient matrix of amount is constituted is designated asWillIn all figures
As the set that the sparse coefficient matrix of characteristic vector is constituted is designated asWillIn
All characteristics of image vectors sparse coefficient matrix constitute set be designated asWherein,WithDimension K × 1, min () is to take minimum value function, symbol " | | | |F" be ask for matrix not Luo Beini crow this
(Frobeniu)-norm sign.
2. _ 4, according to the W obtained in training stage procedure constructionL, estimate LtestIn each sub-block picture quality arrow
Amount, by LtestIn the picture quality vector of t-th of sub-block be designated as And according in training stage process structure
Make obtained WR, estimate RtestIn each sub-block picture quality vector, by RtestIn t-th of sub-block picture quality arrow
Amount is designated as According to the W obtained in training stage procedure constructionC, estimate CtestIn each sub-block image
Quality vector, by CtestIn the picture quality vector of t-th of sub-block be designated as Wherein,With
Dimension be 6 × 1.
2. L _ 5, is calculatedtestPicture quality objective evaluation predicted value, be designated as qL,And calculate
RtestPicture quality objective evaluation predicted value, be designated as qR,Calculate CtestPicture quality objective comment
Valency predicted value, is designated as qC,Wherein,Represent LtestIn t-th of sub-block in all pixels points
The standard deviation of pixel value,Represent RtestIn t-th of sub-block in all pixels point pixel value standard deviation,Represent
CtestIn t-th of sub-block in all pixels point pixel value standard deviation, symbol " | | | |1" it is the 1- norms for asking for matrix
Symbol.
2. _ 6, basisWithIn
The sparse coefficient matrix of all characteristics of image vectors, obtains Ltest、RtestAnd CtestRespective weight coefficient, correspondence is designated as WithThen S is calculatedtestPicture quality objective evaluation predicted value, be designated as Q,
In this particular embodiment, step 2. _ 6 inWithAcquisition process be:
A1, calculatingIn all characteristics of image vectors sparse coefficient matrix histogram, note
For
And calculateIn all characteristics of image vectors sparse coefficient matrix histogram, note
For
CalculateIn all characteristics of image vectors sparse coefficient matrix histogram, be designated as
Wherein, 1≤g≤K, P are representedIn Nogata node of graph sum, also represent
In Nogata node of graph sum, also representIn Nogata node of graph sum, P=is taken in the present embodiment
50, log () are represented with 2 logarithmic functions bottom of for, 1≤p≤P,RepresentIn belong to g-th of dictionary
The probability density function of component,RepresentIn belong to the probability density function of g-th of dictionary component,RepresentIn belong to the probability density function of g-th of dictionary component,RepresentIn belong to the probability density function of p-th of Nogata node of graph,RepresentIn belong to
The probability density function of p Nogata node of graph,RepresentIn belong to the probability of p-th of Nogata node of graph
Density function,RepresentIn g-th of dictionary component it is quantified after coefficient value,RepresentIn g-th of word
Coefficient value after allusion quotation component is quantified,RepresentIn g-th of dictionary component it is quantified after coefficient value,
A2, calculatingWithIn all characteristics of image vectors it is sparse
The joint histogram of coefficient matrix, is designated as Wherein, P is also representedIn it is straight
The sum of square node of graph, 1≤p1≤ P, 1≤p2≤ P,RepresentWith
In belong to the joint probability density function of g-th of dictionary component,RepresentIn belong to pth1It is individual
Nogata node of graph and pth2The joint probability density function of individual Nogata node of graph,
A3, calculating Ltest、RtestAnd CtestRespective independent entropy, correspondence is designated as H (Ltest)、H(Rtest) and H (Ctest),Then count
Calculate LtestAnd RtestCombination entropy, be designated as H (Ltest,Rtest),
A4, calculating Ltest、RtestAnd CtestRespective weight coefficient, is corresponded toWith
Here, evaluating storehouse I and LIVE stereo image quality using University Of Ningbo's stereo-picture storehouse, LIVE stereo image qualities
Storehouse II is evaluated to comment with mean subjective to analyze the picture quality objective evaluation predicted value for the distortion stereo-picture that the present embodiment is obtained
The correlation divided between difference.Take University Of Ningbo's stereo-picture storehouse by 12 undistorted stereo-pictures in different distortion levels
60 width distortion stereo-pictures in the case of 60 width distortion stereo-pictures, JPEG2000 compressions in the case of JPEG compression, Gaussian mode
60 width distortion stereo-pictures in the case of 60 width distortion stereo-pictures and white Gaussian noise in the case of paste.Take LIVE stereo-pictures
80 width distortions in the I of quality evaluation storehouse by 20 undistorted stereo-pictures in the case of the JPEG compression of different distortion levels are three-dimensional
Image, JPEG2000 compression in the case of 80 width distortion stereo-pictures, 45 width distortion stereo-pictures in the case of Gaussian Blur and
80 width distortion stereo-pictures in the case of white Gaussian noise.LIVE stereo image qualities are taken to evaluate undistorted vertical by 8 in the II of storehouse
Body image 72 width distortion stereo-pictures in the case of the JPEG compression of different distortion levels, JPEG2000 compression in the case of 72
72 width distortions in the case of 72 width distortion stereo-pictures and white Gaussian noise in the case of width distortion stereo-picture, Gaussian Blur are stood
Body image.Here, commonly use objective parameter by the use of assess image quality evaluating method 2 and be used as evaluation index, i.e., non-linear time
Pearson correlation coefficient (Pearson linear correlation coefficient, PLCC) under the conditions of returning,
Spearman coefficient correlations (Spearman rank order correlation coefficient, SROCC), PLCC reflections
The accuracy of the objective evaluation result of distortion stereo-picture, SROCC reflects its monotonicity.
University Of Ningbo's stereo-picture storehouse, LIVE stereo image qualities are calculated using the inventive method to evaluate storehouse I and LIVE and stand
The picture quality objective evaluation predicted value of every width distortion stereo-picture in body image quality evaluation storehouse II, recycles existing master
Appearance quality evaluation method obtains University Of Ningbo's stereo-picture storehouse, LIVE stereo image qualities and evaluates storehouse I and LIVE stereo-picture matter
Amount evaluates the mean subjective scoring difference of every width distortion stereo-picture in the II of storehouse.Obtained distortion will be calculated by the inventive method
The picture quality objective evaluation predicted value of stereo-picture does five parameter Logistic function nonlinear fittings, PLCC and SROCC values
It is higher, illustrate that the correlation between objective evaluation result and mean subjective scoring difference is better.Table 1 is given using present invention side
Pearson phases between the picture quality objective evaluation predicted value for the distortion stereo-picture that method is obtained and mean subjective scoring difference
Relation number, table 2 gives the picture quality objective evaluation predicted value of the distortion stereo-picture obtained using the inventive method with putting down
Spearman coefficient correlations between equal subjective scoring difference.As can be seen from Table 1 and Table 2, obtained using the inventive method
Distortion stereo-picture picture quality objective evaluation predicted value and mean subjective scoring difference between correlation be it is very high,
Show that the result of objective evaluation result and human eye subjective perception is more consistent, it is sufficient to illustrate the validity of the inventive method.
The picture quality objective evaluation predicted value and mean subjective for the distortion stereo-picture that table 1 is obtained using the inventive method
Pearson correlation coefficient between scoring difference compares
The picture quality objective evaluation predicted value and mean subjective for the distortion stereo-picture that table 2 is obtained using the inventive method
Spearman coefficient correlations between scoring difference compare