CN104282019A - Blind image quality evaluation method based on natural scene statistics and perceived quality propagation - Google Patents

Abstract

The invention provides a blind image quality evaluation method based on natural scene statistics and perceived quality propagation. The method includes the steps of calculating the fields of experts (FoE) gradient response values of a test image and a large number of undistorted natural images, conducting statistics on the histogram distribution of the response values, calculating the KL divergence of the histogram distribution of the response value of the test image and the histogram distribution of the response values of the undistorted natural images to obtain the absolute distortion degree of the test image, extracting the quality perception characteristics from the test image and marked distorted images, finding N marked images most similar to the test image according to the chi-square distances between the characteristics, conducting weighted summation on the quality grades of the marked images to obtain the relative distortion degree of the test image, and obtaining the final prediction image quality grade by combining the two predicted grades. Compared with an existing representativeness reference-free image quality evaluation method, the method is simple and efficient, and a user does not need to mark a large number of samples through manual work.

Description

Based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated

Technical field

The present invention relates to image processing techniques, particularly perception visual signal treatment technology.

Background technology

Efficient image perception quality evaluating method is then the gordian technique in multimedia service quality monitoring field.At present, reliable image quality evaluating method is mainly full reference and weak reference type.These methods require to access undistorted former figure information completely.But in the middle of many applied environments, this requirement often cannot meet.

Blind image (non-reference picture) quality evaluating method only needs the information of distorted image self and its perceived quality measurable.The black box projection model that existing blind image quality evaluating method has the learning method of supervision direct training image characteristic sum perceived quality to give a mark often through supporting vector recurrence etc. is for image quality estimation.In order to ensure the robustness of model, these methods need a large amount of handmarking's images for training.Meanwhile, because its black box projects, these methods cannot relation between clear Description Image characteristic sum perceived quality.

Summary of the invention

Technical matters to be solved by this invention is, provides the blind reference image quality appraisement method of the relation between a kind of energy Description Image characteristic sum quality score.

The present invention for solving the problems of the technologies described above adopted technical scheme is, based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, comprises the following steps:

Step 1) natural scene statistics:

Calculate expert field FoE (Fields of Experts) the gradient response of test pattern and all undistorted images, and statistics obtains the FoE gradient response histogram distribution P of test pattern respectively _dand the FoE response histogram distribution P of i-th undistorted image _u(i), i=1,2 ..., K, K are undistorted image sum, calculate the relative entropy KL divergence of test pattern and undistorted image $Q_{\mathrm{NSS}}=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{P}_d{\log }_2\left(\frac{P_d}{P_u\left(i\right)}\right);$

Step 2) perceived quality propagation:

1-1: image texture characteristic is extracted to test pattern, overall Gradient Features and the boundary intensity feature based on down-sampling; The extracting method of the boundary intensity feature of described down-sampling is: carry out 1/8 down-sampling to image, after down-sampling the boundary strength value of each point of image by it in vertical and horizontal direction the maximal value of gradient represent, again to boundary strength value a little carry out statistics with histogram, the histogram after normalization is the boundary intensity feature based on down-sampling;

1-2: the feature cards side distance D calculating test pattern and the distorted image to have marked, the described distorted image marked is carried out by the mode of handmarking the distorted image that picture quality gives a mark; wherein F _i ^qand F _i ^rrepresent i-th proper vector of test pattern respectively, mark i-th proper vector of undistorted image, i={1,2,3} be correspondence image textural characteristics respectively, overall Gradient Features and the boundary intensity feature based on down-sampling;

1-3: mark distorted image to little to large select progressively top n according to feature cards side distance D; And calculate respective weight w according to this N number of feature cards side distance D _n, d _nrepresent test pattern and the little feature cards side's distance having marked undistorted image to n-th of large select progressively;

1-4: utilize weight w _nto N number of image quality score DMOS having marked distorted image _nbe weighted the prediction mark Q that summation obtains test pattern _pQP, $Q_{\mathrm{PQP}}=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n\·{\mathrm{DMOS}}_n;$

Step 3) relative entropy KL divergence Q is set _nSSand prediction mark Q _pQPweight parameter, with reference to relative entropy KL divergence Q _nSSand prediction mark Q _pQPobtain final prediction marking Q.

The present invention calculates expert field FoE gradient response to test pattern and a large amount of undistorted natural image, the response histogram distribution of statistical test image and all undistorted images respectively.Again by the KL divergence of both calculating distribution, we can obtain the absolute distortion level of test pattern.Secondly, we to test pattern and the distorted image extraction quality Perception Features that marked, and according to the card side's distance between feature, find the N number of marking image the most similar to test pattern.By the quality score weighted sum by these marking images, the relative distortion level of test pattern can be obtained.Finally, final predicted picture massfraction can be obtained by first two steps prediction marking is grouped together.

The invention has the beneficial effects as follows, compared with in existing representative non-reference picture quality appraisement method, the method is simply efficient, and without the need to a large amount of handmarking's sample.

Accompanying drawing explanation

Fig. 1: block schematic illustration of the present invention.

Embodiment

For effectively carrying out quality assessment to non-reference picture, the present invention is made up of three steps: natural scene statistic procedure, perceived quality propagation steps, comprehensively to give a mark step.Wherein, natural scene adds up the statistical discrepancy by compare test image and a large amount of undistorted natural image, obtains the assessment of absolute distortion information.And mass propagation is by propagating to test pattern similarly by the quality score of portion markings image, the assessment of relative distortion information can be obtained.Finally, by two module prediction marking are combined to obtain final prediction marking.

The present embodiment realizes on matlab2009b software platform, specifically as shown in Figure 1:

The FoE gradient response of step one, calculating test pattern and selected undistorted image, and the histogram distribution both adding up respectively.Allow P _drepresent the response distribution of test pattern, P _urepresent the response distribution of whole undistorted image.Then the quality score of natural scene statistical module can be expressed as both KL divergences, namely

Difference is distributed with as test pattern distortion then responds with the FoE gradient of undistorted image, when test pattern is fuzzy, then the FoE gradient response distribution of test pattern is flat compared with undistorted image, and when test pattern has noise, then its FoE gradient response distribution there will be folder peak.

Step 2, perceived quality are propagated and are formed primarily of following three steps:

1st step: feature test pattern being extracted to mass-sensitive, comprise image texture characteristic, overall situation Gradient Features and the boundary intensity DSBS feature based on down-sampling, wherein image texture characteristic can pass through SFTA (Segmentation-based Fractal Texture Analysis) feature instantiation, and overall Gradient Features is by GIST feature instantiation.For the extraction of the boundary intensity feature based on down-sampling, first 1/8 down-sampling is carried out to image, after sampling the boundary intensity of each point of image by it in vertical and horizontal direction the maximal value of gradient represent.Then, to boundary strength value a little carry out statistics with histogram, the histogram after normalization is DSBS proper vector, and DSBS feature is for reflecting the blocking effect after compression of images.

2nd step: allow F _i ^qand F _i ^ri-th proper vector of expression test pattern and marking image, d (F _i ^q, F _i ^r) card side's distance of both expressions, then test pattern and the total characteristic distance of marking image can be expressed as wherein i={1,2,3} be corresponding SFTA, GIST and DSBS feature respectively.

3rd step: find the 5 width marking image making D minimum, and calculate respective weight w according to their characteristic distance _n:

d _nrepresent test pattern and the little feature cards side's distance having marked undistorted image to n-th of large select progressively;

Then, the prediction of this module marking can be expressed as the weighted sum of the massfraction DMOS of marking image, and the span of massfraction DMOS 0 to 100,0 represents best here, and 100 representatives are the poorest:

$Q_{\mathrm{PQP}}=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n\·{\mathrm{DMOS}}_n$

Step 3, by being combined by first two steps quality score, final prediction marking can be obtained

$Q=Q_{\mathrm{NSS}}^{\mathrm{\γ}}\·Q_{\mathrm{PQP}}^{1-\mathrm{\γ}}$

Wherein γ is the weight parameter of two modules, and here, we are set to 0.2.

Claims (5)

1., based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, it is characterized in that, comprise the following steps:

Step 1) natural scene statistics:

Calculate the expert field FoE gradient response of test pattern and all undistorted images, and statistics obtains the FoE gradient response histogram distribution P of test pattern respectively _dand the FoE response histogram distribution P of i-th undistorted image _u(i), i=1,2 ..., K, K are undistorted image sum, calculate the relative entropy KL divergence of test pattern and undistorted image $Q_{\mathrm{NSS}}=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{P}_d{\log }_2\left(\frac{P_d}{P_u\left(i\right)}\right);$

Step 2) perceived quality propagation:

1-1: image texture characteristic is extracted to test pattern, overall Gradient Features and the boundary intensity feature based on down-sampling; The extracting method of the boundary intensity feature of described down-sampling is: carry out 1/8 down-sampling to image, after down-sampling the boundary strength value of each point of image by it in vertical and horizontal direction the maximal value of gradient represent, again to boundary strength value a little carry out statistics with histogram, the histogram after normalization is the boundary intensity feature based on down-sampling;

1-2: the feature cards side distance D calculating test pattern and the distorted image to have marked, the described distorted image marked is carried out by the mode of handmarking the distorted image that picture quality gives a mark; wherein F _i ^qand F _i ^rrepresent i-th proper vector of test pattern respectively, mark i-th proper vector of undistorted image, i={1,2,3} be correspondence image textural characteristics respectively, overall Gradient Features and the boundary intensity feature based on down-sampling;

1-3: mark distorted image to little to large select progressively top n according to feature cards side distance D; And calculate respective weight w according to this N number of feature cards side distance D _n, d _nrepresent test pattern and the little feature cards side's distance having marked undistorted image to n-th of large select progressively;

1-4: utilize weight w _nto N number of image quality score DMOS having marked distorted image _nbe weighted the prediction mark Q that summation obtains test pattern _pQP, $Q_{\mathrm{PQP}}=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n\·{\mathrm{DMOS}}_n;$

Step 3) relative entropy KL divergence Q is set _nSSand prediction mark Q _pQPweight parameter, with reference to relative entropy KL divergence Q _nSSand prediction mark Q _pQPobtain final prediction marking Q.

2. as claimed in claim 1 based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, it is characterized in that, described prediction marking Q is finally γ is weight parameter.

3., as claimed in claim 2 based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, it is characterized in that, γ=0.2.

4., as claimed in claim 1 based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, it is characterized in that, described down-sampling is 1/8 down-sampling.

5., as claimed in claim 1 based on the blind image quality evaluating method that natural scene statistics and perceived quality are propagated, it is characterized in that, N=5.