CN103325113B - Partial reference type image quality evaluating method and device - Google Patents

Abstract

The invention discloses a kind of partial reference type image quality evaluating method and device, the method comprising the steps of: carry out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively; Normalization conversion is separated to obtain corresponding separation normalization conversion coefficient with distorted image wavelet coefficient separately to reference picture; Obtain the statistical independence between adjacent separation normalization coefficient subband, and using this independence as characteristic information; Feature based information generation unit divides reference image quality evaluation index.Present invention utilizes the rule of adjacent DNT coefficients statistics dependent change during image fault, be extracted the statistical independence feature of reflection image fault degree, construct the objective evaluation Index Formula without the need to fitting parameter, it is more objective to make the evaluation of picture quality.

Description

Partial reference type image quality evaluating method and device

Technical field

The present invention relates to technical field of image processing, particularly relate to a kind of partial reference type image quality evaluating method and device.

Background technology

Image quality evaluation (IQA:ImageQualityAssessment) is devoted to study the perceived quality how using objective indicator evaluation map picture, makes it consistent with image subjective scores height.The work of image quality evaluation algorithm design highly significant, and has wide practical use.Such as, in video quality monitoring field, the cost that contracting workforce carries out monitoring is very high, and reliability is low, and good image quality evaluation algorithm can reduce costs, and raises the efficiency.

In general, the image quality evaluation algorithm of three quasi-representatives is had.According to realization degree from the easier to the more advanced, respectively: full reference type (FR:Full-Reference), partial reference type (RR:Reduced-Reference) and without reference type (NR:No-Reference) IQA.If original image and distorted image are all available, FRIQA utilizes the full detail of this two width image to compare, and calculates its difference.RRIQA method utilizes the partial information of reference picture (usually occurring with a series of RR characteristic formp) predicted picture quality degradation degree, is generally applicable to the application scenario that only can obtain a small amount of statistical information of original image, such as image transfer through networks.NRIQA is the challenging quality assessment task of most, because do not have any information of original image can be for reference, be applicable to the situation that cannot obtain original image information, such as picture quality monitoring.

Such as, Fig. 1 shows the logic diagram of existing partial reference type image QA system, and as shown in Figure 1, the RR feature extracted from original image is transmitted through attached channels, compare with feature that distorted image extracts at receiving end, obtain the quality evaluation index of distorted image.

For performance evaluation, in recent years, some research organizations, according to the relevant criterion of International Telecommunications Union (ITU), establish several more common picture quality subjective assessment database, mainly comprise: the A57 database of Cornell Univ USA's visual communication development in laboratory; U.S.'s oklahoma state university calculates the CSIQ(CategoricalImageQualityDatabase of perception and picture quality development in laboratory) database; The IVC database of the Institute of Technology of Nantes, France university exploitation; Texas,U.S university Austin LIVE(LaboratoryforImageandVideoEngineering) the LIVE database of development in laboratory; Japanese fuji university MICT(MediaInformationandCommunicationTechnology) the MICT database of development in laboratory; And the TID2008 database that Tampere, Finland university and Aero-Space university of Ukraine develop.

Each database substantially all includes common degenerated form as Gauss's additive white noise, Gaussian Blur, high frequency noise, JPEG compression artefacts, JPEG2000 compression artefacts etc., and the subjective scores of each distorted image (MOS:MeanOpinionScore).

According to the guidance standard of ISO (International Standards Organization) video quality expert group (VQEG:VideoQualityExpertsGroup), the predicted value of objective algorithm to subjective quality exists certain non-linear, first should remove this non-linear factor, and then carry out relevance verification.

Remove nonlinear functional form and have multiple choices, such as fitting of a polynomial or Logistic return.Most researchists adopt five parametrical nonlinearity Logistic models of Restricted Linear condition, and LMSE method carries out matching to data, and formula is as follows:

Quality(x)=β ₁logistic(β ₂,(x-β ₃))+β ₄x+β ₅（1）

$\log \mathrm{istic}(\mathrm{\τ},x)=\frac{1}{2}-\frac{1}{1+\exp \left(\mathrm{\τx}\right)}---\left(2\right)$

Wherein, x represents objective RRIQA algorithm predicts value, β ₁, β ₂, β ₃, β ₄, β ₅represent logistic model fitting parameter.

The performance of the objective algorithm of image quality evaluation is generally verified from the accuracy of prediction, monotonicity and consistance three aspects.

In the accuracy of prediction, the difference between the predicted value of objective algorithm and subjective scoring should be the smaller the better.Linearly dependent coefficient CC(LinearCorrelationCoefficient therebetween can be used) and root-mean-square error RMSE(RootMeanSquareError) weigh: for given objective algorithm, CC value is higher, RMSE value is lower, shows that the accuracy that it is predicted is higher.

$\mathrm{CC}=\frac{{\mathrm{\Σ}}_i(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{{\mathrm{\Σ}}_i{(X_i-\stackrel{\‾}{X})}^2\·\sqrt{{\mathrm{\Σ}}_i{(Y_i-\stackrel{\‾}{Y})}^2}}---\left(3\right)$

$\mathrm{RMSE}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(X_i-Y_i)}^2}---\left(4\right)$

Wherein, the objective predicted data of picture quality after X, Y difference representative image quality subjective evaluation experimental data is non-linear with removal, i represents data sequence number, N representative sample number.

For the monotonicity of prediction, the predicted value of objective algorithm should increase and decrease with the increase and decrease of subjective scoring.Can with Spearman rank correlation coefficient SROCC(SpearmanRankOrderCorrelationCoefficient) to measure: SROCC value is higher, and the monotonicity of objective algorithm predicts is better.

$\mathrm{SROCC}=1-\frac{6{\mathrm{\Σ}}_i{D}_i^2}{N(N^2-1)}---\left(5\right)$

Wherein, the difference of the inter-stages such as D representative is the absolute value of difference between subjective experiment data and objective predicted data herein.

For the consistance of prediction, the performance of objective algorithm on test set should be similar to the performance on training set, and namely the Generalization Capability of algorithm is wanted better.Out-of-bounds point rate OR(OutlierRatio can be used) weigh.So-called out-of-bounds point, may be defined as definitely or relative prediction residual exceedes the point of certain threshold value.OR value is lower, and the consistance representing objective algorithm predicts is better.

Except above statistical indicator, by drawing subjective scores scatter diagram corresponding to objective predicted value, observe the performance that its shape also can investigate objective algorithm intuitively.In scatter diagram, usually using the objective algorithm predicts value after nonlinear fitting as horizontal ordinate, using subjective experiment score value as ordinate, to every image to be evaluated, scatter diagram can obtain a point.Meanwhile, by Logistic Function Fitting Drawing of Curve on scatter diagram, each loose some distance Logistic Function Fitting curve is more concentrated, illustrates that algorithm performance is better.

Most of early stage partial reference type image quality evaluation algorithm for certain distortion type, such as image blurring, JPEG blocky effect, JPEG2000 ringing effect etc.Get up for the RRIQA algorithm of general type of distortion in recent years.Wherein, natural image statistics (NSS:NaturalSceneStatistics) model is main model, is accepted and use by most people.The statistical distribution situation of what the existing RRIQA method based on natural image statistical property was investigated usually is linear decomposition coefficient, such as small echo or discrete cosine transform.But some other characteristics of image similar with optic element responding to height are perhaps more useful.

In order to reflect the Nonlinear Mechanism of biological vision system better, a kind of Nonlinear decomposition method, is namely separated normalization conversion (DNT:Divisivenormalizationtransform) and is suggested ^[1].This conversion can carry out efficient coding to natural sign well, and significantly reduces higher order statistical dependence.At present, only having a small amount of research work DNT thought to be introduced RRIQA, such as Li and Wang utilizes this conversion to obtain the class Gaussian distribution feature of subband ^[2], and be applied to partial reference type image quality assessment, further will introduce this below.

Fig. 2 a shows original image.Fig. 2 b shows the schematic diagram by the wavelet coefficient subband obtained after carrying out linear decomposition to the original image in Fig. 2 a, available linear decomposition mode has a lot, here steerable pyramid is selected to decompose (Steerablepyramiddecomposition), because this conversion can set scale parameter and the direction number of wavelet decomposition easily and flexibly, thus obtain the partial descriptions of image on yardstick, direction and locus.Fig. 2 c shows the schematic diagram by carrying out the rear DNT coefficient subband obtained of DNT conversion to each wavelet coefficient subband in Fig. 2 b, and DNT conversion is based upon on to the linear decomposition basis of image, if y represents wavelet coefficient, then and normalization coefficient , wherein p is for being separated normalized factor (p>0), represents the energy of coefficient y at yardstick, direction and adjacent coefficient collection spatially.

The method calculating p has many, adopts the method based on Gauss's yardstick mixing (GSM:Gaussianscalemixtures) model herein.Length is the condition that the random vector Y of N is defined as GSM,

It can be represented as the product of two isolated components, that is:

Wherein, represent probability with distribution, U is covariance is C _uzero-mean gaussian random vector, z be called mixing multiplier random scalar.

From formula (6), the probability that GSM model describes random vector has identical covariance (C _u) and the mixing of Gaussian random vector of different zoom ratio (z).If the probability density of mixing multiplier is p _z(z), then the probability density of Y is:

$p_Y\left(Y\right)=\∫\frac{1}{{\left[2\mathrm{\π}\right]}^{\frac{N}{2}}{\left|z^2{C}_U\right|}^{1/2}}\exp (-\frac{Y^T{C}_U^{-1}Y}{2z^2})p_z\left(z\right)\mathrm{dz}---\left(7\right)$

Wherein, random vector Y is defined as space adjacent coefficient (eight neighborhood) of certain point in subband, and same position place yardstick adjacent coefficient, direction adjacent coefficient set, verified its meets GSM model.

Herein in application, be simplify to calculate, regulation z all gets fixed value in each position, thus Y is reduced to covariance is z ²c _uzero-mean gaussian vector.Contrast equation (6) and DNT definition, expect that normalized factor p obtains by calculating adjacent coefficient vector Y (known by the view data) estimation to multiplier z very naturally.Coefficient vector Y is traveled through each wavelet coefficient subband along moving window, will the normalized factor p with spatial position change be produced.Only operation is normalized to window center coefficient yc, obtains normalized coefficient wherein, it is the estimation of z.Utilize maximum likelihood estimate, can obtain:

$\hat{z}=\arg \underset{z}{\max }\left\{\log p\left(Y\right|z)\right\}$

$=\arg \underset{z}{\min }\{N\log z+Y^T{C}_U^{-1}Y/2z^2\}$

$=\sqrt{Y^T{C}_U^{-1}Y/N}---\left(8\right)$

Wherein, covariance matrix C _u=E [UU ^t] can be estimated obtain by image wavelet coefficient of dissociation, N is the length of coefficient vector Y, i.e. the number of neighborhood (yardstick, direction, space adjacent) wavelet coefficient.

In extraction characteristic aspect, above-mentioned Li and Wang has added up the coefficient histogram of each DNT subband of natural image, finds its approximate Gaussian distributed.In contrast thereto, the subband DNT coefficient histogram of distorted image is different with type of distortion, deviates from Gaussian curve by different way.

Specifically, Fig. 3 a shows original image, and Fig. 3 b shows the Gaussian noise image of the original image in Fig. 3 a, and Fig. 3 c shows the Gaussian Blur image of the original image in Fig. 3 a, and Fig. 3 d shows the jpeg image of the original image in Fig. 3 a.Herein, can using the original image in Fig. 3 a as reference image.Correspondingly, Fig. 4 a shows the DNT sub-band coefficients histogram of image in Fig. 3 a and reference picture, Fig. 4 b shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 b and reference picture, Fig. 4 c shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 c and reference picture, Fig. 4 d shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 d and reference picture, wherein, solid line is the DNT sub-band coefficients histogram of distorted image, and dotted line is the DNT sub-band coefficients histogram of reference picture.

Utilize KLD distance (Kullback-LeiblerDistance, KLD) of reference picture DNT sub-band coefficients histogram and distorted image DNT sub-band coefficients histogram curve, the feature of representative image objective quality can be extracted be defined as follows:

Wherein, p (x) and q (x) represents original and subband DNT coefficient histogram that is distorted image respectively, d (p q) the KLD distance of p (x) and q (x) is represented.

From formula (9), calculate need p (x) and q (x) known.For RRIQA application, receiving end cannot obtain complete original image subband DNT coefficient histogram p (x).Fortunately, from Fig. 3 (a), p (x) and Gaussian curve p _mx () is closely similar:

$p_m\left(x\right)=\frac{1}{\sqrt{2\mathrm{\π}}\mathrm{\σ}}\exp (-\frac{x^2}{2{\mathrm{\σ}}^2})---\left(10\right)$

Wherein, σ represents the variance of Gaussian function.

At transmitting terminal, p _mthe KLD distance of (x) and p (x) computable:

Meanwhile, if parameter σ is sent by attached channels by transmitting terminal, so just can reconstruction of function p at receiving end _m(x), thus can p be calculated _mthe KLD distance of (x) and q (x) :

Due to with are all computable (or known), thus can obtain further estimated value objective indicator as image quality evaluation:

Present discussion evaluated error.By formula (11) and (12) for people's formula (13), obtain:

Thus with between evaluated error be:

Visible, work as p _m(x) and p (x) closely time, the result of above formula levels off to zero, can ignore.And for natural image, this characteristic ubiquity (as shown in Fig. 3 a and 4a).Therefore, the estimated distance found is effective in theory.

Except li and Wang finds that following 3 kinds of distance feature are also helpful for raising RR image quality index assessed for performance:

d _σ=|σ _o-σ _d|（16）

Wherein, σ _oand σ _drepresent the variance of original image and distorted image DNT coefficient respectively.

d _k=|k _o-k _d|（17）

Wherein, k _oand k _drepresent the kurtosis of original image and distorted image DNT coefficient respectively.

d _s=|s _o-s _d|（18）

Wherein, s _oand s _drepresent the degree of bias of original image and distorted image DNT coefficient respectively.

Based on above four features, next need integration structure image quality evaluation criterion computing formula, be defined as follows:

Wherein, α, beta, gamma and δ represent joint four distance feature coefficient separately respectively, can be obtained, D by the data learning training in conventional images storehouse _bandrepresent the image quality evaluation values of a certain selected DNT subband.

Finally, then by the quality assessment value D of each subband _bandsummation, obtains the Objective image quality evaluation index D of view picture distorted image.

Existing scheme is when extracting feature, and done an important hypothesis, that is: the DNT coefficient histogram of natural image is Gaussian distributed.But discovery when we carry out emulation experiment to the reference picture in more public image storehouses, has quite a few natural image and disobeys this rule.Therefore, when carrying out matching with Gaussian curve, will with larger error.In addition, prior art have selected four kinds of distance feature, and wherein the selection of the first has clear and definite theory support; Then the selection gist of three kinds is all not mentioned.Finally, these four kinds dissimilar distance feature are integrated by existing scheme, have used four different coefficients.These coefficients are comparatively strong to the dependence of data-base content, and use different image libraries instead, the coefficient obtained is also different, are difficult to determine optimum and the most pervasive coefficient.

Summary of the invention

The technical problem to be solved in the present invention is to adopt different coefficients for needing when carrying out partial reference type image quality assessment in prior art, and this coefficient cannot provide the defect of objective appraisal because of comparatively strong to the dependence of data-base content, provide a kind of partial reference type image quality evaluating method and device.

The technical solution adopted for the present invention to solve the technical problems is: provide a kind of partial reference type image quality evaluating method, comprise step:

S100, respectively wavelet decomposition is carried out to obtain respective wavelet coefficient to reference picture and distorted image;

S200, normalization conversion is separated to obtain corresponding separation normalization conversion coefficient to reference picture with distorted image wavelet coefficient separately;

S300, obtain statistical independence between adjacent separation normalization coefficient subband, and using statistical independence as characteristic information;

S400, feature based information generation unit divide reference image quality evaluation index.

In the partial reference type image quality evaluating method of the foundation embodiment of the present invention, in the step s 100,

Adopt steerable pyramid to decompose and respectively wavelet decomposition is carried out to obtain respective wavelet coefficient to reference picture and distorted image.

According in the partial reference type image quality evaluating method of the embodiment of the present invention, in step S300,

Statistical independence is obtained by the mutual information calculated between adjacent separation normalization coefficient subband.

According in the partial reference type image quality evaluating method of the embodiment of the present invention, in step S300,

Adjacent three aspects adjacent with position, direction adjacent from yardstick obtain the adjacent statistical independence being separated normalization conversion coefficient subband using as characteristic information respectively.

According in the partial reference type image quality evaluating method of the embodiment of the present invention, in step S400,

According to following formula generating portion reference image quality evaluation index:

Wherein, D is the city distance of distorted image and reference picture, for arbitrary characteristic information of distorted image, for with the characteristic information of corresponding reference picture.

Present invention also offers a kind of partial reference type image quality evaluation device, comprising:

Wavelet decomposition module, for carrying out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively;

Be separated normalization conversion module, for being separated normalization conversion to obtain corresponding separation normalization conversion coefficient to reference picture with distorted image wavelet coefficient separately;

Characteristic information extracting module, for obtaining the statistical independence between adjacent separation normalization coefficient subband, and using statistical independence as characteristic information;

Image quality evaluation index generation module, divides reference image quality evaluation index for feature based information generation unit.

In the partial reference type image quality evaluation device of the foundation embodiment of the present invention, wavelet decomposition module user adopts steerable pyramid decomposition to carry out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively.

In the partial reference type image quality evaluation device of the foundation embodiment of the present invention, characteristic information extracting module is used for obtaining statistical independence by the mutual information calculated between adjacent separation normalization coefficient subband.

According in the partial reference type image quality evaluation device of the embodiment of the present invention, characteristic information extracting module is used for adjacent three aspects adjacent with position, direction adjacent from yardstick respectively and extracts the adjacent statistical independence being separated normalization conversion coefficient subband using as characteristic information.

In the partial reference type image quality evaluation device of the foundation embodiment of the present invention, image quality evaluation index generation module is used for according to following formula generating portion reference image quality evaluation index:

Wherein, D is the city distance of distorted image and reference picture, for arbitrary characteristic information of distorted image, for with the characteristic information of corresponding reference picture.

The beneficial effect that the present invention produces is: the rule that present invention utilizes adjacent DNT coefficients statistics dependent change during image fault, be extracted the statistical independence feature of reflection image fault degree, construct the objective evaluation Index Formula without the need to fitting parameter, it is more objective to make the evaluation of picture quality.

Accompanying drawing explanation

Below in conjunction with drawings and Examples, the invention will be further described, in accompanying drawing:

Fig. 1 shows the logic diagram of existing partial reference type image QA system;

Fig. 2 a shows the schematic diagram of original image;

Fig. 2 b shows the schematic diagram by the wavelet coefficient subband obtained after carrying out linear decomposition to the original image in Fig. 2 a;

Fig. 2 c shows the schematic diagram by carrying out the rear DNT coefficient subband obtained of DNT conversion to each wavelet coefficient subband in Fig. 2 b;

Fig. 3 a shows original image;

Fig. 3 b shows the Gaussian noise image of the original image in Fig. 3 a;

Fig. 3 c shows the Gaussian Blur image of the original image in Fig. 3 a;

Fig. 3 d shows the jpeg image of the original image in Fig. 3 a;

Fig. 4 a shows the DNT sub-band coefficients histogram of image in Fig. 3 a and reference picture;

Fig. 4 b shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 b and reference picture;

Fig. 4 c shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 c and reference picture;

Fig. 4 d shows the DNT sub-band coefficients histogram of distorted image in Fig. 3 d and reference picture;

Fig. 5 shows the logic diagram of the partial reference type image quality evaluation device according to the embodiment of the present invention;

Fig. 6 shows the detailed logic block diagram of the partial reference type image quality evaluation device in Fig. 5;

Fig. 7 a shows the schematic diagram decomposing the rear wavelet coefficient subband obtained by carrying out steerable pyramid to the original image in Fig. 2 a;

The schematic diagram of the DNT subband that Fig. 7 b obtains after showing and carrying out being separated normalization conversion to the wavelet coefficient in Fig. 7 a;

Fig. 8 a and 8b respectively illustrates the adjacent DNT subband adjacent with direction of yardstick, i.e. the joint distribution figure of the subband dn2.1 that is adjacent respectively of subband dn1.1 and dn1.2 coefficient;

Fig. 9 a-9d respectively illustrates the joint histogram of the DNT coefficient subband of adjacent 1 pixel of level in JPEG2000, JPEG, white noise, Gaussian Blur image;

Figure 10 a-10e is the conditional histograms of JPEG distorted image in distortion level enhancing situation gradually in Fig. 9 b;

Figure 11 shows the Venn diagram of descriptive information entropy mutual relationship;

Figure 12 shows the process flow diagram of the partial reference type image quality evaluating method according to the embodiment of the present invention.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

Fig. 5 shows the logic diagram of the partial reference type image quality evaluation device according to the embodiment of the present invention, and this device comprises wavelet decomposition module 100, is separated normalization conversion module 200, characteristic information extracting module 300 and image quality evaluation index generation module 400.Wherein, wavelet decomposition module 100 receives reference picture and the distorted image of input, and carries out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively.Be separated normalization conversion module 200 pairs of reference pictures and be separated the separation normalization conversion coefficient that normalization conversion obtains correspondence with distorted image wavelet coefficient separately.Characteristic information extracting module 300 can obtain the statistical independence between adjacent separation normalization coefficient subband, and using this statistical independence as characteristic information.Image quality evaluation index generation module 400 feature based information generation unit divides reference image quality evaluation index.

Specifically, Fig. 6 shows the detailed logic block diagram of the partial reference type image quality evaluation device in Fig. 5, as shown in Figure 6, wavelet decomposition module 100 comprises the first wavelet decomposition unit 110 being positioned at transmitting terminal and the second wavelet decomposition unit 120 being positioned at receiving end; Separation normalization conversion module 200 comprises the first separation normalization converter unit 210 being positioned at transmitting terminal and is separated normalization converter unit 220 be positioned at receiving end second; Characteristic information extracting module 300 comprises the fisrt feature information extraction unit 310 being positioned at transmitting terminal and the second feature information extraction unit 320 being positioned at receiving end.The course of work of partial reference type image quality evaluation device will be described in detail below based on Fig. 6.

First, the reference picture of the first wavelet decomposition unit 110 receiving end/sending end input, also can be referred to as original image, and carry out wavelet decomposition to obtain its wavelet coefficient to reference picture.Herein, available linear decomposition mode has a lot, steerable pyramid such as can be selected to decompose, because this conversion can set scale parameter and the direction number of wavelet decomposition easily and flexibly, thus obtain the partial descriptions of image on yardstick, direction and locus.Still using the image in Fig. 2 a as original image, Fig. 7 a shows the schematic diagram decomposing the rear wavelet coefficient subband obtained by carrying out steerable pyramid to the original image in Fig. 2 a.

Correspondingly, the second wavelet decomposition unit 120 receives distorted image at receiving end, and this distorted image is sent from transmitting terminal by reference picture, is formed after arriving receiving end after distorting channel transmission.The course of work of the second wavelet decomposition unit 120 is identical with the first wavelet decomposition unit 110, carries out wavelet decomposition to obtain its wavelet coefficient to distorted image, and the wavelet coefficient and Fig. 7 a that decompose acquisition are similar.Herein, steerable pyramid can be selected to decompose distorted image.

First is separated the wavelet coefficient that normalization converter unit 210 receives the reference picture that the first wavelet decomposition unit 110 exports, and carries out separation normalization conversion to obtain its separation normalization conversion coefficient (i.e. DNT) coefficient to this wavelet coefficient.Such as, Fig. 7 b shows the schematic diagram carrying out being separated the rear DNT subband obtained of normalization conversion to the wavelet coefficient in Fig. 7 a.Wherein, also show center coefficient and corresponding yardstick neighbours, direction neighbours and spatial neighbors.

Correspondingly, second is separated normalization converter unit 220 and first, and to be separated the course of work of normalization converter unit 210 identical, it receives the wavelet coefficient of the distorted image that the second wavelet decomposition unit 120 exports, and carries out separation normalization conversion to obtain its DNT coefficient to this wavelet coefficient.DNT coefficient subband and the subband shown in Fig. 7 b of distorted image are similar.

First system feature information extraction unit 310 obtains the statistical independence between adjacent separation normalization coefficient subband for reference picture, and the characteristic information using this statistical independence as above-mentioned reference picture; Second feature information extraction unit 320 obtains the statistical independence between adjacent separation normalization coefficient subband for distorted image, and the characteristic information using this statistical independence as above-mentioned distorted image.Because fisrt feature information extraction unit 310 is identical with the course of work of second feature information extraction unit 320, the discussion therefore will be suitable for both simultaneously.

Statistical independence between natural image DNT coefficient was studied in great detail by early stage researchist.In order to the statistical independence between same distorted image DNT coefficient clearly contrasts, we are still for example bright here.Fig. 8 a and 8b respectively illustrates the adjacent DNT subband adjacent with direction of yardstick, i.e. the joint distribution of the subband dn2.1 that is adjacent respectively of subband dn1.1 and dn1.2 coefficient.Wherein, each row value of joint histogram has separately carried out convergent-divergent to fill completely whole dynamic range.

Diagramatic curve in Fig. 8 a and Gaussian curve closely similar, average and the variance of conditional histograms function histo (YX) are approximately constant, with X dn2.1 value have nothing to do.Herein, (X, Y) represents the DNT coefficient after a pair adjacent quantification.Obviously, through DNT conversion, single order and second-order statistics independence all obtain reinforcement.For the DNT coefficient that direction is adjacent, the variance of conditional histograms function still remains unchanged, but average changes with the change of the X in dn1.2, shows to there is correlativity between adjacent direction coefficient.

For the joint distribution of all the other neighbouring relations DNT coefficients, the same universal law existed as shown in figure 8 a, that is: to remain constant constant for the variance of conditional histograms, shows the statistical independence that DNT space is potential.

For RRIQA problem, the statistical independence Changing Pattern of distorted image is even more important.Fig. 9 a shows the joint histogram of the DNT coefficient subband of adjacent 1 pixel of level in JPEG2000 image; Fig. 9 b shows the joint histogram of the DNT coefficient subband of adjacent 1 pixel of level in jpeg image; Fig. 9 c shows the joint histogram of the DNT coefficient subband of adjacent 1 pixel of level in white noise image; Fig. 9 d shows the joint histogram of the DNT coefficient subband of adjacent 1 pixel of level in Gaussian Blur image.The reference picture of a series of different type of distortion in above-mentioned figure and distortion level (from top to bottom increasing progressively) image is all from the natural image in Fig. 2 a.

In order to understand the Changing Pattern of statistical independence better, taken out by the conditional histograms of JPEG distorted image in Fig. 9 b and compare in Figure 10 a-10e, wherein, distortion level strengthens gradually by Figure 10 a.Clearly, along with increasing progressively of distortion level, histogrammic average and variance are no longer fixed as constant, and the shape of each curve also more and more departs from Gaussian curve.The diversity factor of each conditional histograms curve shape promotes, and also reveal that the progressively forfeiture of statistical independence degree.

As can be seen from Figure 10 a, 10b and 10d, for distorted image, statistical independence no longer exists.Distortion level is more serious, and conditional histograms variance heterogeneity is higher.Show that the destruction of picture structure result in the decline of statistical independence.Seem in Figure 10 c to have occurred one " counter-example ", along with distortion level increases, statistical independence progressively strengthens.Consider that white noise itself is exactly completely independently with picture material, add more white noises in the picture, also just mean and introduce more statistical iteration compositions, the phenomenon observed in Figure 10 c is also logical.Namely the change degree of statistical independence is consistent with distortion level, with the form being directly proportional or being inversely proportional to.

Although show only the statistical independence Changing Pattern on adjacent level direction herein.In fact, to in the experiment of other neighboring condition, comprise adjacent yardstick, adjacent direction, adjacent position (different directions and distance), and the neighboring condition of whole DNT subbands of other image, in DNT territory, the change of statistical independence presents its ubiquity, and the degree of change almost becomes to quantize corresponding relation with distortion level, thus, adopt the index parameter of statistical independence as quantized image distortion level of adjacent DNT coefficient subband in the present invention, i.e. characteristic information.

For partial reference type image quality evaluating method, need an objective indicator that can reflect statistical independence degree.If stochastic variable (X, Y) represents the DNT coefficient after a pair quantification, (x _i, y _j), i, j=1,2 ... N, the possible value of representative (X, Y), when meeting the following conditions, X and Y is defined as separate:

P(X=x _i,Y=y _j)=P(X=x _i)P(Y=y _j)(21)

Wherein, P represents probability.

More than definition illustrates how to judge that whether two stochastic variables are separate.But two stochastic variables not only degree that interdepends immediately still have no way of learning.For measuring the degree that statistics relies on further, preferably, can select mutual information (MI:MutualInformation), it is defined as follows:

$\begin{array}{c}I(X;Y)=H\left(X\right)+H\left(Y\right)-H(X,Y)---\left(22\right)\\ H\left(X\right)=-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}P{(X=x_i)}_i\log P(X=x_i)---\left(23\right)\\ H\left(Y\right)=-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}P(Y=y_j)\log P(Y=y_j)---\left(24\right)\\ H(X,Y)=-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}P(X=x_i,Y=y_j)\log P(X=x_i,Y=y_j)---\left(25\right)\end{array}$

Wherein, I (X; Y) be the mutual information of X and Y, H (X) and H (Y) represents the information entropy of X and Y respectively, and H (X, Y) is the combination entropy of (X, Y).

Figure 11 shows the Venn diagram (VennDiagram) of descriptive information entropy mutual relationship.As seen from Figure 11, X and Y is more independent, I (X; Y) less.In extreme circumstances, when X and Y is completely independent, I (X; Y) 0 is equaled.

Therefore, be in preferred embodiment in the present invention, characteristic information extracting module 300 obtains the statistical independence degree of this adjacent DNT subband by the mutual information calculated between adjacent DNT coefficient subband.Specifically, the mutual information that fisrt feature information extraction unit 310 obtains its adjacent DNT subband for reference picture is used as its statistical independence, and the mutual information that second feature information extraction unit 320 obtains its adjacent DNT subband for distorted image is used as its statistical independence.

Preferably, can adjacent from yardstick, direction is adjacent and the statistical independence of adjacent DNT coefficient subband is extracted in adjacent three aspects, position.That is, characteristic information can carry out the statistical independence of the adjacent DNT coefficient subband of the statistical independence of the adjacent DNT coefficient subband in the statistical independence of the adjacent DNT coefficient subband of self scale, direction and position.Specifically, as shown in Figure 7b, adjacent sub-bands is defined as subband pair adjacent on yardstick, direction and position.Such as, for subband dn1.3, its adjacent scale subbands is dn2.3; Adjacent directional subband is dn1.2 and dn1.4; Adjacent position subband lays respectively at 0 ,-45, on-90 and-135 degree directions.For convenience of understanding, wherein one group of adjacent position subband is to being marked by solid box in fig .7b.

In an embodiment of the present invention, when selecting the wavelet decomposition in three yardstick four directions, the characteristic information (4 groups of adjacent direction × 3 yardsticks) that the adjacent characteristic information (direction, 2 groups of adjacent yardstick × 4) of 8 yardsticks, 12 directions are adjacent can be extracted, the characteristic information (4 adjacent space pixel × 12 subbands) that 48 positions are adjacent, adds up to 68 characteristic informations.

The characteristic information extracted for reference picture is sent to the image quality evaluation index generation module 400 being positioned at receiving end by the fisrt feature information extraction unit 310 being positioned at transmitting terminal through attached channels, characteristic information is directly sent to image quality evaluation index generation module 400, with synthetic image quality evaluation index for distorted image extraction unit by the second feature information extraction unit 320 being positioned at receiving end.

In an embodiment of the present invention, although the quantity of characteristic information is many compared with prior art, it extracts type is but consistent, and the method for extraction is also identical, namely all obtains according to the statistics degree of dependence of adjacent DNT coefficient subband.Therefore, this programme feature is weighted without the need to arranging coefficient respectively, only the difference of distorted image and each character pair of reference picture need be sued for peace, can obtain distorted image total quality metric.Be defined as follows:

Wherein, D is the city distance of distorted image and reference picture, represent a certain eigenwert of distorted image, represent reference picture characteristic of correspondence value.Above-mentioned city distance D can be adopted as image quality evaluation index.

As can be seen from the above, present invention utilizes the rule of adjacent DNT coefficients statistics dependent change during image fault, be extracted the statistical independence feature of reflection image fault degree, construct the objective evaluation Index Formula without the need to fitting parameter.Utilize existing common data database data to carry out emulation experiment discovery, the RRIQA index that this programme extracts is simple and effective, can reflect the distortion level of distorted image preferably.Therefore can be applicable to partial reference type image quality assessment problem, can be used as the objective performance index of picture quality, instruct a series of related application problems such as Internet Transmission.

Figure 12 shows the process flow diagram of the partial reference type image quality evaluating method according to the embodiment of the present invention, above-described partial reference type image quality evaluation device can be adopted to implement the method, therefore, the above description about partial reference type image quality evaluation device is partly or entirely quoted herein.By step, this partial reference type image quality evaluating method will be described below.

S100, respectively wavelet decomposition is carried out to obtain respective wavelet coefficient to reference picture and distorted image.Preferably, steerable pyramid can be adopted to decompose and respectively wavelet decomposition to be carried out to obtain respective wavelet coefficient to reference picture and distorted image.

S200, normalization conversion is separated to obtain corresponding separation normalization conversion coefficient to reference picture with distorted image wavelet coefficient separately.

S300, obtain statistical independence between adjacent separation normalization coefficient subband, and using this statistical independence as characteristic information.Preferably, by calculating mutual information between adjacent separation normalization coefficient subband to obtain statistical independence.And the adjacent statistical independence being separated normalization conversion coefficient subband can be extracted in adjacent three aspects adjacent with position, direction adjacent from yardstick respectively.S400, feature based information generation unit divide reference image quality evaluation index.Preferably, can according to formula (26) generating portion reference image quality evaluation index.

Should be understood that, for those of ordinary skills, can be improved according to the above description or convert, and all these improve and convert the protection domain that all should belong to claims of the present invention.

List of references:

[1]O.Schwartz,andE.P.Simoncelli,“NaturalSignalStatisticsandSensoryGainControl,”NatureNeuroscience,vol.4,no.8,pp.819–825,2001;

[2]Q.Li,andZ.Wang,“Reduced-ReferenceImageQualityAssessmentUsingDivisiveNormalization-BasedImageRepresentation,”IEEEJournalofSelectedTopicsinSignalProcessing,vol.3,no.2,pp.202–211,2009。

Claims (4)

1. a partial reference type image quality evaluating method, is characterized in that, comprises step:

S100, respectively wavelet decomposition is carried out to obtain respective wavelet coefficient to reference picture and distorted image;

S200, normalization conversion is separated to obtain corresponding separation normalization conversion coefficient to described reference picture with described distorted image described wavelet coefficient separately;

S300, obtain statistical independence between adjacent separation normalization coefficient subband, and using described statistical independence as characteristic information;

S400, based on described characteristic information generating portion reference image quality evaluation index;

In described step S100, adopt steerable pyramid to decompose and respectively wavelet decomposition is carried out to obtain respective wavelet coefficient to reference picture and distorted image; Selecting steerable pyramid to decompose, by setting scale parameter and the direction number of wavelet decomposition, obtaining the partial descriptions of image on yardstick, direction and locus;

In described step S300, obtain described statistical independence by the mutual information calculated between adjacent separation normalization coefficient subband;

In described step S300, adjacent three aspects adjacent with position, direction adjacent from yardstick obtain the adjacent described statistical independence being separated normalization conversion coefficient subband using as described characteristic information respectively.

2. partial reference type image quality evaluating method according to claim 1, is characterized in that, in described step S400,

Described partial reference type image quality evaluation index is generated according to following formula:

Wherein, D is the city distance of described distorted image and described reference picture, for arbitrary characteristic information of described distorted image, for with the characteristic information of corresponding reference picture.

3. a partial reference type image quality evaluation device, is characterized in that, comprising:

Wavelet decomposition module, for carrying out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively;

Be separated normalization conversion module, for being separated normalization conversion to obtain corresponding separation normalization conversion coefficient to described reference picture with described distorted image described wavelet coefficient separately;

Characteristic information extracting module, for obtaining the statistical independence between adjacent separation normalization coefficient subband, and using described statistical independence as characteristic information;

Image quality evaluation index generation module, for based on described characteristic information generating portion reference image quality evaluation index;

Described wavelet decomposition module user adopts steerable pyramid decomposition to carry out wavelet decomposition to obtain respective wavelet coefficient to reference picture and distorted image respectively; Selecting steerable pyramid to decompose, by setting scale parameter and the direction number of wavelet decomposition, obtaining the partial descriptions of image on yardstick, direction and locus;

Described characteristic information extracting module is used for obtaining described statistical independence by the mutual information calculated between adjacent separation normalization coefficient subband;

Described characteristic information extracting module is used for adjacent three aspects adjacent with position, direction adjacent from yardstick respectively and extracts the adjacent described statistical independence being separated normalization conversion coefficient subband using as described characteristic information.

4. partial reference type image quality evaluation device according to claim 3, is characterized in that, described image quality evaluation index generation module is used for generating described partial reference type image quality evaluation index according to following formula:

Wherein, D is the city distance of described distorted image and described reference picture, for arbitrary characteristic information of described distorted image, for with the characteristic information of corresponding reference picture.