CN103366378A - Reference-free type image quality evaluation method based on shape consistency of condition histogram - Google Patents

Abstract

The invention discloses a reference-free type image quality evaluation method based on the shape consistency of a condition histogram. The reference-free type image quality evaluation method comprises the following steps: S1: carrying out nonlinear processing on a distorted image; S2: carrying out feature extraction on the image subjected to the nonlinear processing and extracting a combined histogram and a condition histogram curve histo (Y/Xi); solving a condition histogram average value; S3: utilizing the condition histogram curve histo (Y/Xi) and the condition histogram average value to solve a quality evaluation index D for judging the image quality. The quality evaluation index D obtained by using the method disclosed by the invention is simple and practical and the distortion degree of the distorted image can be effectively reflected; compared with a data training method, according to the method disclosed by invention, the obtained evaluation index has better objectivity and stability.

Description

Based on the conforming no-reference image quality evaluation method of condition histogram shape

Technical field

The present invention relates to image processing field, particularly relate to the image quality evaluation field, relate in particular to a kind of based on the conforming no-reference image quality evaluation method of condition histogram shape.

Background technology

Picture quality be evaluation map as the leading indicator of processing system for video and algorithm performance, therefore extremely important to the research of image quality evaluation (Image Quality Assessment, IQA) method and have a widely prospect.In the processing procedures such as image denoising, image recovery, figure image intensifying, the IQA index can be used as the foundation of each algorithm performance comparison, parameter selection; Image Coding with the field such as communicate by letter, the IQA index can be used for instructing whole compression of images, transmission, receiving course, and assessment algorithms of different and system performance; In Internet Transmission quality monitoring field, the IQA index can replace manually-operated, realizes automatic real time monitoring; In addition, IQA has potential practical value at numerous ambits such as image processing algorithm optimization, living things feature recognition, Medical Image Compression.

Present digital picture quality evaluation (IQA) research can be divided into subjective evaluation method and method for objectively evaluating.

Subjective evaluation method is to utilize the subjective assessment of eye-observation to come the evaluation map image quality, specifically, under the conditions such as certain image source, display device and environment, two width of cloth pictures are provided simultaneously for the beholder, wherein a width of cloth is original image, another width of cloth is distorted image, and beholder wherein comprises ordinary people and image professional and layman.Score data according to each beholder counts the data such as average, standard deviation, 95% confidence interval, with the evaluation map image quality.Because the people is the ultimate recipient of image vision information, so utilizing subjective experiment to come the evaluation map image quality is the most accurate and effective method, yet the experimental data that evaluation method needs is very large, and being difficult for embedding automated system adds up, cause this evaluation method to waste time and energy, and be subject to subjective factor and watch the impact of environment, be unfavorable in practice widespread use.

Method for objectively evaluating comprises full reference type (Full-Reference, FR) image quality evaluation (hereinafter to be referred as: FR-IQA), partial reference type (Reduced-Reference, RR) image quality evaluation (hereinafter to be referred as: RR-IQA) and without reference type (No-Reference, NR) image quality evaluation (hereinafter to be referred as: NR-IQA).Wherein, blind image quality evaluation (Blind IQA) is named again in the no-reference image quality evaluation.FR-IQA is fully known as the original image information of estimating the distorted image quality references and without any distortion.RR-IQA only utilizes the partial information of original image to come the visually-perceptible quality of distortion estimator image.NR-IQA is a kind of any information that does not need original image, directly distorted image is carried out Appraising Methods.FR-IQA and RR-IQA all or part of dependence original and undistorted image as a reference, and in realization the whole even partial information of reference picture to be difficult to the cost that obtains or obtain higher, and hdr video need not reference picture and also can make an appraisal to picture quality, so the tool Research Significance of NR-IQA.The NR-IQA algorithm can be divided into for single distortion with for the algorithm of general distortion, and wherein single distortion comprises the algorithm that the distortions such as white Gaussian noise, Gaussian Blur, JPEG, JPEG2000 or channel fading to image are estimated; And general distortion is the algorithm that above-mentioned single type of distortion is carried out mixing evaluation, has widely applicable surface.

For evaluation map image quality better, in recent years, some research organizations have set up several more common picture quality subjective assessment databases according to the relevant criterion of International Telecommunications Union (ITU), mainly comprise: the A57 database of Cornell Univ USA's visual communication development in laboratory; U.S.'s oklahoma state university calculates (Categorical Image Quality Database) database of perception and picture quality development in laboratory; The IVC database of the Institute of Technology of Nantes, France university exploitation; Austin branch school LIVE(Laboratory for Image and Video Engineering of texas,U.S university) the LIVE database of development in laboratory; The MICT(Media Information and Communication Technology of Japanese fuji university) the MICT database of development in laboratory; And the TID2008 database of Tampere, Finland university and the exploitation of Ukraine Aero-Space university.Each database has substantially all comprised common deteriroation of image quality type such as Gauss's additive white noise, Gaussian Blur, high frequency noise, JPEG compression artefacts, JPEG2000 compression artefacts, and the data such as the subjective score of each distorted image (Mean Opinion Score, MOS).

According to (the Video Quality Experts Group of ISO (International Standards Organization) video quality expert group, VQEG) guidance standard, objective algorithm exists certain non-linear to the predicted value of picture quality subjective assessment, therefore, at first should remove this non-linear factor when utilizing objective algorithm to image quality evaluation, and then carry out the correlativity checking.The functional form of non-linear removal has multiple choices, such as fitting of a polynomial or Logistic recurrence etc.

The correlativity checking refers to the objective Algorithm Performance of image quality evaluation is verified from accuracy, monotonicity and the consistance of prediction.

The accuracy of prediction refers to that the difference of the predicted value of objective algorithm and subjective scoring value is more little more accurate.Usually use the linearly dependent coefficient CC(Linear Correlation Coefficient between the two) and root-mean-square error RMSE(Root Mean Square Error) weigh: for given objective algorithm, the CC value is higher, the RMSE value is lower, shows that the accuracy of its prediction is higher.

$\mathrm{CC}=\frac{{\mathrm{\Σ}}_i(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{\sqrt{{\mathrm{\Σ}}_i{(X_i-\stackrel{\‾}{X})}^2}\·\sqrt{{\mathrm{\Σ}}_i{(Y_i-\stackrel{\‾}{Y})}^2}}$ （1.1）

$\mathrm{RMSE}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(X_i{-Y}_i\right)}^2}$ （1.2）

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

The monotonicity of prediction refers to that the predicted value of objective algorithm should increase and decrease with the increase and decrease of subjective scoring value.Usually use Spearman rank correlation coefficient SROCC(Spearman Rank Order Correlation Coefficient) to measure: the SROCC value is higher, and the monotonicity of objective algorithm predicts is better.

$\mathrm{SROCC}=1-\frac{6{\mathrm{\Σ}}_i{D}_i^2}{N(N^2-1)}$ （1.3）

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

The consistance of prediction refers to that objective algorithm should be similar to the performance on training set in the performance on the test set, and namely the Generalization Capability of algorithm is wanted better.Usually use out-of-bounds point rate OR(Outlier Ratio) weigh.So-called out-of-bounds point may be defined as the point that absolute or relative prediction residual surpasses certain threshold value.The OR value is lower, and the consistance that represents objective algorithm predicts is better.

Except above index, can also by drawing the subjective score value and the corresponding scatter diagram of objective predicted value of getting, observe its shape and also can investigate intuitively objective Algorithm Performance.In scatter diagram, usually with the objective algorithm predicts value behind the nonlinear fitting as horizontal ordinate, as ordinate, to every image to be evaluated, on scatter diagram, can both obtain a point with the subjective experiment score value.Simultaneously, on scatter diagram, each loose some distance L ogistic Function Fitting curve is more concentrated, illustrates that algorithm performance is better with Logistic Function Fitting Drawing of Curve.

In to the no-reference image quality evaluation procedure, utilize statistical property (the Natural scene statistics of natural scene, NSS) extracting evaluating characteristic is practices well, because the statistical property of natural scene can be changed losing true time, as long as be a series of IQA features with this change portrayal artificially, just can identify certain distortion.A kind of based on statistical independence (STAtistical INDependence, STAIND) no-reference image quality evaluation method (Y.Chu, X.Mou, W.Hong, and Z.Ji, " A novel no-reference image quality assessment metric based on statistical independence; " Proceedings of2012IEEE Visual Communications and Image Processing, pp.1 – 6, article Nov.2012.) discloses under a kind of NSS framework and based on adjacent DNT coefficients statistics independence no-reference image quality have been carried out Appraising Methods.

Adjacent separation normalization conversion (the Divisive normalization transform of image after correlative study discovery natural image and the distortion thereof, hereinafter to be referred as DNT) have the statistics dependence between the coefficient, that is: in the natural image each adjacent DNT coefficient be approximate statistical independently; And for distorted image, along with the aggravation of distortion level, statistical independence also changes thereupon, or direct proportion or inverse proportion variation.Therefore correlation technique utilizes the mutual information that extracts in the adjacent DNT coefficient joint histogram to portray above-mentioned statistical independence as the core of no-reference image quality evaluation method.The step of no-reference image quality evaluation method comprises distorted image is carried out feature extraction in the open correlation technique of Fig. 1, again feature and the training image that extracts is input to the enterprising row operation of regression model together, obtains the picture quality scoring.Characteristic extraction procedure wherein comprises image is carried out wavelet decomposition, DNT conversion, calculates joint histogram, extracts four steps of mutual information.Fig. 2 shows distorted image is carried out the process that coefficient that wavelet decomposition obtains carries out the DNT conversion, after the DNT conversion, obtain distorted image DNT coefficient, comprise that the DNT coefficient is at yardstick, direction and locational adjacent coefficient collection, if establish stochastic variable (X, Y) the DNT coefficient after a pair of quantification of representative, (x _i, y _j), i, j=1,2 ... N, the possible value of representative (X, Y), when meeting the following conditions, X and Y are defined as separate:

P(X＝x _i,Y＝y _j)＝P(X＝x _i)P(Y＝y _j) (1.4)

Wherein, P represents probability.

Concentrate the joint histogram that extracts under the different neighbouring relations (it is adjacent with yardstick to comprise that direction is adjacent, the position is adjacent) from above-mentioned adjacent coefficient, calculate again its mutual information (mutual information, MI), utilize mutual information to judge statistical independence.Mutual information is defined as follows:

I(X;Y)＝H(X)+H(Y)-H(X,Y) (1.5)

$H\left(X\right)=-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}P{(X=x_i)}_i\log P(X=x_i)$ (1.6)

$H\left(Y\right)=-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}P(Y=y_j)\log P(Y=y_j)$ (1.7)

$H(X,Y)=-\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}P(X=x_i,Y=y_j)\log P\left(X=x_i,Y=y_i\right)$ (1.8)

Wherein, I (X; Y) be the mutual information of X and Y, H (X) and H (Y) represent respectively the information entropy of X and Y, and H (X, Y) is the combination entropy of (X, Y).Facts have proved that almost for all type of distortion, the mutual information value is all with the distortion level monotone variation, as for type of distortion such as JPEG2000 and JPEG, mutual information is monotone increasing along with the aggravation of distortion level; For white Gaussian noise, the mutual information value is monotone decreasing along with the aggravation of distortion level; For Gaussian Blur, mutual information is just along with distortion level aggravates and monotone increasing when distortion level is not serious, when distortion level is serious, descend on the contrary, therefore it also not exclusively satisfies the monotonicity requirement, therefore the evaluation index of utilizing mutual information to make is only followed monotonicity in the part situation, can not be suitable for the quality assessment of all type of distortion images.

After determining mutual information, need to determine a specific features extraction scheme, below as an example of Fig. 2 example the specific features extraction scheme in the explanation correlation technique.Fig. 2 shows that distorted image is carried out steerable pyramid to be decomposed and separate the normalization conversion after (being wavelet decomposition) again and obtain DNT and decompose subband, adjacent sub-bands wherein refers to subband adjacent on yardstick, direction and position pair, as for subband dn1.3, its adjacent scale subbands is dn2.3; Adjacent directional subband is dn1.2 and dn1.4; The adjacent position subband lays respectively at 0, on-45 ,-90 and-135 degree directions.Selected among Fig. 2 three yardstick four directions to wavelet decomposition, therefore it can extract 8 adjacent scale features (2 groups of adjacent yardstick * 4 directions), 12 adjacent direction characters (4 groups of adjacent direction * 3 yardsticks), 48 adjacent position features (4 adjacent space pixel * 12 subbands) add up to 68 features.

After extracting the correlated characteristic data by above-mentioned feature extraction scheme, it is merged into overall objective, with the quality of predicted distortion image accurately and effectively.In the routine techniques, support vector machine (Support Vector Machine, SVM) is applied in the design of NR-IQA method as forecast model.The feature that extracts in the distorted image is sent into SVM carry out regression training, obtain Parameters in Regression Model, be used for follow-up image quality evaluation values prediction.Usually select ε-support vector regression (Support vector regression, SVR) and radial basis function (Radial basis function, RBF) as kernel function.In the regression process, the optimum value of Parameters in Regression Model c and g all obtains by training, and other parameter uses as default, value can be referring to list of references (C.Chang in detail, and C.Lin, " LIBSVM:a library for support vector machines " Http:// www.csie.ntu.edu.tw/～cjlin/libsvm.).

Adopt the method according to the regression parameter of training image feature acquisition nonlinear model in the such scheme, this method need to be carried out complicated data training, and evaluation result is subject to the impact of picture material and training strategy, the objectivity deficiency.In addition, Parameters in Regression Model will change with the variation of training dataset, and optimized parameter determines that by iteration tests the value on different type of distortion, disparate databases is all different, thereby the stability of its predicted distortion picture quality is not enough, is not suitable for actual NR-IQA system.

Summary of the invention

The technical problem to be solved in the present invention is estimated learning method for relevant NR-IQA and is existed the data training complicated, and evaluation result stability is not enough; And Parameters in Regression Model is that the variation with training dataset changes, and causes the uncertain deficiency of Parameters in Regression Model value, provide a kind of need not to train based on the conforming no-reference image quality evaluation method of condition histogram shape.

The present invention is achieved by the following technical programs: a kind of based on the conforming no-reference image quality evaluation method of condition histogram shape, comprise the steps:

S1: distorted image is carried out Nonlinear Processing;

S2: will carry out feature extraction through the image of Nonlinear Processing, and extract joint histogram and condition histogram curve histo (Y|X _i), and solving condition histogram average

S3: utilize condition histogram curve histo (Y|X _i) and condition histogram average

Solve to judge the quality evaluation index D of picture quality.

In the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, comprise that also distorted image is carried out wavelet decomposition obtains multiple dimensioned multidirectional wavelet decomposition subband before the Nonlinear Processing among the described step S1; Described Nonlinear Processing is that the coefficient of wavelet decomposition with each wavelet decomposition subband separates normalization (DNT) conversion, obtains the process that multiple dimensioned multidirectional DNT decomposes subband.

In the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, feature extraction among the described step S2 is to decompose the DNT that extracts the adjacent position the subband from multiple dimensioned multidirectional DNT to decompose subband, after described adjacent position being decomposed the DNT coefficient discretize of subband, form a series of joint histograms, from joint histogram, extract again N condition histogram curve histo (Y|X _i); Described condition histogram average Find the solution as follows:

$\stackrel{\‾}{\mathrm{histo}\left(Y\right|X)}=\frac{1}{N}\mathrm{\Σhisto}\left(Y\right|X_i),$ X _i＝X ₁,X ₂,...,X _N；

Wherein, histo (Y|X _i) represent X and get X _iThe time the condition histogram, N represents the curve number of condition histogram curve group.

In the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, described step S3 may further comprise the steps:

S301: utilize condition histogram curve histo (Y|X _i) and condition histogram average

Solve the conforming data characteristics f of condition histogram shape of evaluation map image quality _iThe conforming data characteristics f of condition histogram shape _iMethod for solving is as follows: $f_i=\frac{1}{N}\mathrm{\Σd}\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\‾}{\mathrm{histo}\left(Y\right|X)}\right)$

Wherein, f _iRepresentative condition histogram curve group shape consistance feature; I is data number, maximal value n=positional number * scale parameter of i * direction number;

Represent curve histo (Y|X _i) and condition histogram average

The KLD distance.

S302: utilize the conforming data characteristics f of condition histogram shape _iBy

Construct multi-C vector

Described n refers to the dimension of multi-C vector;

S303: utilize the conforming data characteristics f of condition histogram shape _iBy

Solve multi-C vector

L ₁The model vector length

Wherein n refers to the dimension of multi-C vector, and i is data number;

S304: the multi-C vector of taking from right image

L ₁The model vector length Mean value in some databases is the cluster centre value

S305: pass through multi-C vector

L ₁The model vector length With the cluster centre value

Distance calculate the quality evaluation index D that judges picture quality, wherein

In the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, described in the step S301

Method for solving as follows:

$d\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\‾}{\mathrm{histo}\left(Y\right|X)}\right)=\∫\mathrm{histo}\left(Y\right|X_i)\log \frac{\mathrm{histo}\left(Y\right|X_i)}{\stackrel{\‾}{\mathrm{histo}\left(Y\right|X)}}\mathrm{dt}.$

In the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, the some databases among the step S304 comprise LIVE, CSIQ and TID2008 database.

The present invention compared with prior art has following advantage: the condition histogram shape conformance law that passes through to analyze natural image DNT conversion among the present invention, the characteristics of recycling natural image feature clustering, calculate stable cluster numerical value, and definite no-reference image quality evaluation index D, D is simple and practical for this index, need not training, and can effectively reflect the distortion level of distorted image.Condition histogram among the present invention is to extract from the decomposition subband of adjacent position, and only consider the condition histogram of adjacent position when condition histogram shape consistance analyzed, and need not to consider other neighbouring relations, can avoid like this variety classes neighbouring relations feature span different, linear fit is brought inconvenience, thereby make the more convenient structure of NR-IQA index.

Description of drawings

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

Fig. 1 is correlation technique is carried out quality assessment to distorted image process flow diagram.

Fig. 2 is that correlation technique is carried out the synoptic diagram that steerable pyramid decomposes and separate normalization conversion and feature extraction to distorted image.

Fig. 3 is the process flow diagram based on the conforming no-reference image quality evaluation method of condition histogram shape of the present invention.

Fig. 4 is that the present invention complies with the described synoptic diagram that distorted image is carried out steerable pyramid decomposition and separation normalization conversion and feature extraction based on the conforming no-reference image quality evaluation method of condition histogram shape.

Fig. 5 extracts the joint histogram that adjacent position DNT decomposes subband from the DNT decomposition subband of Fig. 4.

Fig. 6 is the condition histogram that Fig. 5 (b) extracts.

Fig. 7 is condition histogram and the joint histogram of the adjacent DNT coefficient in natural image position among Fig. 4.

Fig. 8 is adjacent position condition histogram curve group and condition histogram average figure.

Embodiment

Understand for technical characterictic of the present invention, purpose and effect being had more clearly, now contrast accompanying drawing and describe the specific embodiment of the present invention in detail.

Fig. 3 discloses the process flow diagram that the present invention is based on the conforming no-reference image quality evaluation method of condition histogram shape, in conjunction with Fig. 1 as can be known, this method need to be carried out feature extraction to distorted image equally, but in correlation technique, distorted image is carried out wavelet decomposition, the DNT conversion, calculate joint histogram, extract the process of mutual information, extract feature in this method and need to pass through wavelet decomposition, the DNT conversion, calculate joint histogram, extraction conditions histogram and calculate the several steps of shape consistance, the present invention that different from correlation technique is adopts the condition histogram that extracts the adjacent position from joint histogram and calculates its shape consistance as the feature extraction of distorted image out.And the characteristics of image that extracts in the correlation technique is to be input to the SVM regression model to carry out regression training, obtain Parameters in Regression Model, be used for the successive image prediction of quality, and the condition histogram shape consistance that extracts among the present invention is direct and natural image cluster point merges to carry out feature, obtains the picture quality scoring.It is more stable to utilize condition histogram shape consistance feature and natural image cluster centre to collect data among the present invention, can not cause the training of correlation technique factor data complicated, cause the unstable and inaccurate problem of quality evaluation result, and need not to realize image quality evaluation through the training of complexity, simple to operate, the result is accurate.

Below will each step that Fig. 3 mentions be described based on the conforming no-reference image quality evaluation method of condition histogram shape disclosed by the invention, the method comprises the steps:

S1: distorted image is carried out Nonlinear Processing, distorted image is converted into numerical characteristics.Specifically, before to the distorted image Nonlinear Processing, first distorted image is carried out wavelet decomposition, thereby obtain multiple dimensioned multidirectional wavelet decomposition subband, adopt steerable pyramid to decompose (Steerable pyramid decomposition) in the present embodiment distorted image is carried out wavelet decomposition.Nonlinear Processing is exactly that coefficient of wavelet decomposition with each wavelet decomposition subband separates normalization (DNT) conversion, obtains the process that multiple dimensioned multidirectional DNT decomposes subband.The explanation distorted image decomposes the process of separating again normalization (DNT) conversion after (Steerable pyramid decomposition) through steerable pyramid as an example of Fig. 4 example.

Among Fig. 4, distorted image decomposes by steerable pyramid, the coefficient of wavelet decomposition spi.j that obtains decomposing, (wherein i and j represent respectively scale parameter and direction number); Thereby obtain the partial descriptions of distorted image on yardstick, direction and locus.Represent wavelet coefficient if establish y, then normalization coefficient is

Wherein p represents the energy of the adjacent coefficient collection of coefficient y on yardstick, direction and space for separating normalized factor (p＞0).

Separating the normalization operation is to utilize Gauss's yardstick to mix (Gaussian scale mixtures, GSM) model the image that decomposes through steerable pyramid is processed the process that obtains separating normalized factor p.

The condition that the random vector Y that length is N is defined as GSM is that it can be represented as the product of two isolated components, that is:

$Y\stackrel{\·}{=}\mathrm{ZU}$ （2.1）

Wherein,

Represent probability with distributing, U is that covariance is C _UThe zero-mean Gaussian random vector, z is called the at random scalar that mixes multiplier.By following formula as can be known, to describe the probability of random vector be to have identical covariance (C to the GSM model _U) and the mixing of the Gaussian random vector of different zoom ratio (z).

If mixing the probability density of multiplier is p _z(z), then the probability density of Y is:

$p_Y\left(Y\right)=\&Integral;\frac{1}{{\left[2\mathrm{\&pi;}\right]}^{\frac{\mathrm{<figure-callout\; id="2"\; label="N"\; filenames="HDA00003574722900041.png"\; state="\{\{state\}\}">N</figure-callout>}}{2}}{\left|Z^2{C}_U\right|}^{1/2}}\exp (-\frac{Y^T{C}_U^{-1}Y}{2z^2})p_z\left(z\right)\mathrm{dz}$ （2.2）

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

In practice, calculate for simplifying, regulation z all gets fixed value in each position, be z thereby Y is reduced to covariance ²C _UZero-mean Gauss vector.Contrast equation (2.1) and DNT definition expect that very naturally normalized factor p can obtain by the estimation of calculating adjacent coefficient vector Y (known by view data) to multiplier z.Coefficient vector Y is traveled through each wavelet coefficient subband along moving window, can be with the normalized factor p of spatial position change with producing.Only to window center coefficient y _cCarry out the normalization operation, obtain normalized coefficient.

Wherein,

It is the estimation of z.Utilize maximum likelihood estimate, can get:

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

$=\arg \underset{z}{\min }\{N\log z+Y^T{C}_U^{-1}Y/{2z}^2\}$ （2.3）

$=\sqrt{Y^T{C}_U^{-1}Y/N}$

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

For separating normalized factor p.

With coefficient of wavelet decomposition spi.j, (wherein i and j represent respectively scale parameter and direction number) carries out the DNT conversion, be about to coefficient of wavelet decomposition spi.j, (wherein i represents respectively scale parameter and direction number with j) is divided by separating normalized factor p, obtain the coefficient of dissociation dni.j that multiple dimensioned multidirectional DNT decomposes subband, (wherein i and j represent respectively scale parameter and direction number).

S2: will carry out feature extraction through the image of Nonlinear Processing, as shown in Figure 4, in the present embodiment equally to distorted image carry out three yardstick four directions to wavelet decomposition, and extracted the adjacent position subband 0,-45,-90 and-135 spend the subband feature of the adjacent sub-bands on the directions, namely are equivalent to only extract in 68 technical characterictics of correlation technique joint histogram and the condition histogram curve histo (Y|X of 48 adjacent position features (4 adjacent space pixel * 12 subbands) _i), and solving condition histogram average Only propose the adjacent position feature in the scheme and do not consider yardstick and direction character, main cause is that the method only need consider that the adjacent position relationship characteristic can obtain the quality assessment of high accuracy, if scale feature and direction character are all extracted the difficulty that can increase the method processing when calculating, because the data characteristics span of adjacent position and adjacent yardstick, adjacent direction is also inconsistent, process complicated.To decompose the DNT that extracts the adjacent position the subband from multiple dimensioned multidirectional DNT to decompose subband specifically, after described adjacent position being decomposed the DNT coefficient discretize of subband, form a series of joint histograms, from joint histogram, extract again N condition histogram curve histo (Y|X _i); Described condition histogram average

Through type (2.4) is found the solution:

$\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}=\frac{1}{N}\mathrm{\&Sigma;histo}\left(Y\right|X_i),$ X _i＝X ₁,X ₂,...,X _N (2.4)

Wherein, histo (Y|X _i) represent X and get X _iThe time the condition histogram, N represents the curve number of one-dimensional condition histogram curve group.

Fig. 5, Fig. 6 reflection proposes joint histogram and the histogrammic process of condition to distorted image, wherein, Fig. 5 be DNT is decomposed adjacent position in the subband the decomposition subband extraction out, a series of joint histograms that construct.The reference picture of distortion level (the from top to bottom increasing progressively) image of the types such as a series of different type of distortion JPEG2000, JPEG, white noise and Gaussian Blur is all from natural image among Fig. 4 among Fig. 5.This figure has reflected the statistical independence Changing Pattern of distorted image.For showing conveniently, carried out independently respectively convergent-divergent to fill completely whole dynamic range to uniting histogrammic each row value.The change degree of statistical independence is consistent with distortion level, with the form that is directly proportional or is inversely proportional to.The condition histogram functions histo (Y|X that Fig. 6 extracts the condition histogram of JPEG distorted image in (b) among Fig. 5 _i).

Utilize condition histogram functions histo (Y|X _i) solve condition histogram average Condition histogram average

Solution procedure as follows:

$\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}=\frac{1}{N}\mathrm{\&Sigma;histo}\left(Y\right|X_i),$ X _i＝X ₁,X ₂,...,X _N （2.5）

Wherein, histo (Y|X _i) represent that X gets X _iThe time the condition histogram, N represents condition histogram histo (Y|X _i) the curve number of curve group.The condition histogram average here

It is not the index that directly reflects the distorted image shape degree of consistency.The condition histogram average that utilization solves

With condition histogram curve histo (Y|X _i) group can construct adjacent position condition histogram curve group and condition histogram average figure (as shown in Figure 8), although can not directly objectively as the evaluation index of no-reference image quality, still can see intuitively its shape degree of consistency from this figure.

As shown in Figure 7, in the embodiment of the invention, because the joint histogram that extracts according to the feature extraction scheme among the step S2 is 17 * 17 grades two-dimensional discrete value, for avoiding the fringe region impact, only chooses the condition histogram curve group of joint histogram center strip location and calculate.Therefore, X ₁Equal 5, X ₂Equal 6; By that analogy, X _NEqual 13; N equals 9.

Fig. 8 has showed natural image position neighbor histogram curve histo (Y|X shown in Figure 4 _i) (fine rule) and average thereof

(thick line).Among Fig. 8, the distortion level of image is less, and the shape similarity of its whole suite line is higher, curve histo (Y|X _i) apart from average

Value less; The distortion level of image is larger, and the shape consistance of whole suite line is poorer, curve histo (Y|X _i) apart from average

Value larger.

S3: utilize condition histogram curve histo (Y|X _i) and condition histogram average

Solve to judge the quality evaluation index D of picture quality.Specifically, comprise the steps:

S301: utilize condition histogram curve histo (Y|X _i) and condition histogram average Solve the conforming data characteristics f of condition histogram shape of evaluation map image quality _iThe conforming data characteristics f of condition histogram shape _iMethod for solving is as follows:

$f_i=\frac{1}{N}\mathrm{\&Sigma;d}\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}\right)$ （2.6）

Wherein, f _iRepresentative condition histogram curve group shape consistance feature; I is data number, and its maximal value n gets the value of the positional number * scale parameter of wavelet decomposition * direction number;

Represent curve histo (Y|X _i) and condition histogram average

The KLD distance.

Method for solving as follows:

$d\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}\right)=\&Integral;\mathrm{histo}\left(Y\right|X_i)\log \frac{\mathrm{histo}\left(Y\right|X_i)}{\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}}\mathrm{dt}$ （2.7）

In most cases, the conforming data characteristics f of condition histogram shape of through type (2.6) and formula (2.7) picture quality finding the solution out _iTo type distortions such as typical JPEG, JPEG2000, Gaussian Blurs, direct proportion changes with the distortion level aggravation; To white Gaussian noise distortion, f _iInversely proportional variation.Therefore, to general NR-IQA, still can't directly utilize f _iThe quality evaluation index of linear fit distorted image.

S302: utilize the conforming data characteristics f of condition histogram shape _iBy

Construct multi-C vector

Described n refers to the dimension of multi-C vector.Because of in the present embodiment to natural image carry out be three yardstick four directions to wavelet decomposition, therefore

Maximum occurrences=positional number of middle n * scale parameter * direction number, i.e. n=4 * 3 * 4=48 is so can pass through in this scheme $\stackrel{\&RightArrow;}{f_{\mathrm{location}}}\stackrel{\mathrm{\&Delta;}}{=}(f_1,{\mathrm{<figure-callout\; id="2"\; label="f"\; filenames="HDA00003574722900041.png"\; state="\{\{state\}\}">f</figure-callout>}}_2,...,f_{48})$ Solve multi-C vector

S303: utilize the conforming data characteristics f of condition histogram shape _iBy

Solve multi-C vector

L ₁The model vector length

Wherein n refers to the dimension of multi-C vector, and i is data number.The observation of nature image

L ₁The value of model vector length on a plurality of public image databases found Value is highly stable, is approximately constant, presents good Clustering features.

S304: the multi-C vector of taking from right image

L ₁The model vector length

Mean value in some databases is the cluster centre value

Some databases here comprise LIVE, CSIQ and three databases of TID2008.(seeing Table 1)

Table 1LIVE, CSIQ, TID2008 database nature reference picture

Statistical value relatively

By as seen from Table 1, on LIVE, CSIQ and TID2008 three large database concepts, natural image condition histogram shape consistance proper vector

L ₁The model vector length

Variance yields is very little, shows excellent stability, and its mean approximation is constant, therefore gets its mean value on each database As cluster centre, thus the stability of the evaluation result of ensuring the quality of products.Experimental data shows the natural image feature of extracting by the present invention

Can depart from gradually with the aggravation of distortion level cluster centre.

S305: pass through multi-C vector

L ₁The model vector length With the cluster centre value Distance calculate the quality evaluation index D that judges picture quality, wherein

Be worth littlely, illustrate that the quality of its distorted image is better.

With the quality evaluation index D value of picture quality with its predicted value with subjective assessment carried out correlativity verify.Because it is non-linear that the value of the quality evaluation index D that use the present invention obtains exists, therefore carrying out at first will removing non-linear factor before the correlativity checking.Adopt five parametrical nonlinearity Logistic models of Restricted Linear condition in the present embodiment, and the least mean-square error method is carried out match to data:

Quality(x)＝β ₁logistic(β ₂,(x-β ₃))+β ₄x+β ₅ （2.8）

$\log \mathrm{istic}(\mathrm{\&tau;},x)=\frac{1}{2}-\frac{1}{1+\exp \left(\mathrm{\&tau;x}\right)}$ （2.9）

Wherein, x represents objective RR-IQA algorithm predicts value, β ₁, β ₂, β ₃, β ₄, β ₅Represent logistic models fitting parameter.

Remove quality evaluation index D value non-linear by five parametrical nonlinearity Logistic models after, its correlativity index verified to be placed on the open LIVE database according to the NR-IQA index D value that the present invention extracts carry out emulation experiment, by its accuracy of CC value test, its monotonicity of SROCC value test, as shown in table 2 below, illustrate that this NR-IQA index can reflect the distortion level of distorted image well.

The CC of table 2 this programme method on the LIVE database, the SROCC value

Mixing distortion in the table 2 is with the type of distortion such as JPEG2000, JPEG, white Gaussian noise, Gaussian Blur and channel fading are carried out the accuracy (CC) of the evaluation index of mixing evaluation and the value of monotonicity (SROCC), mainly for the no-reference image quality evaluation method of general distortion.Table 2 reflection utilizes image quality evaluation index that this method obtains all having preferably accuracy and monotonicity for JPEG2000, JPEG, white Gaussian noise, Gaussian Blur and the channel fading of natural image and the evaluation index of mixing the type of distortion such as distortion, can estimate multiple type of distortion image, be that its evaluation index has very high accuracy and monotonicity to type of distortion evaluations such as white Gaussian noise and Gaussian Blurs especially.The present invention describes by several specific embodiments, it will be appreciated by those skilled in the art that, without departing from the present invention, can also carry out various conversion and be equal to alternative the present invention.In addition, for particular condition or concrete condition, can make various modifications to the present invention, and not depart from the scope of the present invention.Therefore, the present invention is not limited to disclosed specific embodiment, and should comprise the whole embodiments that fall in the claim scope of the present invention.

Claims (6)

1. one kind based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: comprise the steps:

S1: distorted image is carried out Nonlinear Processing;

S2: will carry out feature extraction through the image of Nonlinear Processing, and extract joint histogram and condition histogram curve histo (Y|X _i), and solving condition histogram average

S3: utilize condition histogram curve histo (Y|X _i) and condition histogram average Solve to judge the quality evaluation index D of picture quality.

2. according to claim 1 based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: comprise that also distorted image is carried out wavelet decomposition obtains multiple dimensioned multidirectional wavelet decomposition subband before the Nonlinear Processing among the described step S1; Described Nonlinear Processing is that the coefficient of wavelet decomposition with each wavelet decomposition subband separates normalization (DNT) conversion, obtains the process that multiple dimensioned multidirectional DNT decomposes subband.

3. according to claim 2 based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: the feature extraction among the described step S2 is to decompose the DNT that extracts the adjacent position the subband from multiple dimensioned multidirectional DNT to decompose subband, after described adjacent position being decomposed the DNT coefficient discretize of subband, form a series of joint histograms, from joint histogram, extract again N condition histogram curve histo (Y|X _i); Described condition histogram average

Find the solution as follows:

$\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}=\frac{1}{N}\mathrm{\&Sigma;histo}\left(Y\right|X_i),$ X _i＝X ₁,X ₂,...,X _N；

Wherein, histo (Y|X _i) represent X and get X _iThe time the condition histogram, N represents the curve number of condition histogram curve group.

4. according to claim 3 based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: described step S3 may further comprise the steps:

S301: utilize condition histogram curve histo (Y|X _i) and condition histogram average

Solve the conforming data characteristics f of condition histogram shape of evaluation map image quality _iThe conforming data characteristics f of condition histogram shape _iMethod for solving is as follows: $f_i=\frac{1}{N}\mathrm{\&Sigma;d}\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}\right)$

Wherein, f _iRepresentative condition histogram curve group shape consistance feature; I is data number, maximal value n=positional number * scale parameter of i * direction number;

Represent curve histo (Y|X _i) and condition histogram average

The KLD distance.

S302: utilize the conforming data characteristics f of condition histogram shape _iBy Construct multi-C vector

Described n refers to the dimension of multi-C vector;

S303: utilize the conforming data characteristics f of condition histogram shape _iBy Solve multi-C vector L ₁The model vector length

Wherein n refers to the dimension of multi-C vector, and i is data number;

S304: the multi-C vector of taking from right image

L ₁The model vector length

Mean value in some databases is the cluster centre value

S305: pass through multi-C vector

L ₁The model vector length

With the cluster centre value

Distance calculate the quality evaluation index D that judges picture quality, wherein

5. according to claim 4 based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: described in the step S301

Method for solving as follows: $d\left(\mathrm{histo}\left(Y\right|X_i)\right|\left|\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}\right)=\&Integral;\mathrm{histo}\left(Y\right|X_i)\log \frac{\mathrm{histo}\left(Y\right|X_i)}{\stackrel{\&OverBar;}{\mathrm{histo}\left(Y\right|X)}}\mathrm{dt}.$

6. according to claim 4 based on the conforming no-reference image quality evaluation method of condition histogram shape, it is characterized in that: the some databases among the step S304 comprise LIVE, CSIQ and TID2008 database.

Images