CN104408736A - Characteristic-similarity-based synthetic face image quality evaluation method - Google Patents

Abstract

The invention discloses a characteristic-similarity-based synthetic face image quality evaluation method, and aims to mainly solve the problem of poor consistency of subjective and objective results when a conventional image quality evaluation method is directly applied to a synthetic face image. An implementation process comprises the following steps: 1) preprocessing a synthetic face image to be tested and a reference image to make the sizes of the synthetic face image to be tested and the reference image consistent; 2) partitioning the synthetic face image and the reference image into blocks; 3) performing quality evaluation on synthetic face image blocks of each image to obtain quality evaluation scores of the image blocks; 4) calculating the quality evaluation score of whole image. Compared with the conventional method, the method has the advantages that the particularity of a synthetic face image structure is taken fully into account, the subjective and objective quality evaluation consistency of the synthetic face image is improved, and the method can be used for criminal suspect identification.

Description

The synthesis human face portrait quality evaluating method of feature based similarity

Technical field

The invention belongs to technical field of image processing, relate to synthesis human face portrait quality evaluating method, can be used for identification.

Background technology

Along with the popularization that information intelligent and face recognition technology are applied, such as public security department suspect recognition system, corporate facility face recognition work-checking machine etc., people more and more pay attention to person identification and authentication techniques.But the situation of loss of learning can be there is in real life, as criminal investigation mechanism chase suspect time, the photo of suspect often cannot directly obtain.Offer help for pursuing and capturing an escaped prisoner, only have artist to describe suspect's portrait of drafting according to eyewitness, and the existing database of public security organ being face picture data storehouse at present, identifies under needing that human face photo-portrait is converted into same expression way.But changing in the process of portrait by human face photo, the quality of the synthesis human face portrait obtained has distortion in various degree, and the quality of synthesizing human face portrait quality will directly affect the accuracy determining suspect.Therefore wish, when using synthesis human face portrait method, the good synthetic method of synthesis human face portrait quality can be selected, so it is most important for raising discrimination effectively to carry out the quality assessment of synthesis human face portrait.

Because current image quality evaluating method research spininess both domestic and external is to the natural image of all kinds of distortion, the mainly fuzzy and noise considered in experimentation, such as gaussian mask, Gaussian noise, salt-pepper noise, high frequency noise, JEPG distortion, JEPG2000 distortion etc.; And the blocking effect that the problem affecting human face portrait quality assessment mainly exists and human face structure, so existing image quality evaluating method to be applied in human face portrait quality assessment aspect effect not satisfactory.

Traditional Objective image quality evaluation method mainly concentrates on research natural image, and it is mainly divided three classes:

One, full-reference image quality evaluating method

Full reference image quality appraisement method be when original image information as a reference carry out image quality evaluation, the thought of these class methods is image quality from the similarity degree of two images, hypothetical reference image is standard picture, if reference picture and image to be evaluated just the same, then similarity degree is 100%; If there is distortion, similarity degree is between 0-l.The method of current existence mainly contains the methods such as PSNR, SSIM, FSIM, IW-SSIM, VIF, VSNR.

Two, partial reference type image quality evaluating method

Partial reference type image quality evaluating method only uses the partial information of original image to carry out assess image quality.The method of current existence mainly contains the methods such as WN-ISM, RRED.

Three, no-reference image quality evaluation method

The common ground of full reference type and partial reference type image quality evaluating method is: they rely on original image as a reference whole or in part.But in fact in many situations all cannot or the undistorted original image of more difficult acquisition, such as imaging system, user terminal etc., front two class evaluation methods just cannot play a role in these cases.In addition, consider human visual system evaluate piece image quality time and without the need for original image as reference, even if the image information received is limited, people also can make evaluation to picture quality very like a cork, and this is a kind of evaluation method without reference type inherently.The main research method existed is the methods such as BIQI, BLIINDS, LBIQ, DIIVINE, BLINDS-II, CORNIA at present.

When said method is applied to the quality assessment of synthesis human face portrait, do not consider the structural difference of human face portrait and natural image and the blocking effect of human face portrait, cause the consistance of quality evaluation result and subjective quality assessment low, effect is not satisfactory.

Summary of the invention

The object of the invention is to for the deficiency of above-mentioned conventional images quality evaluating method in the application of synthesis human face portrait, a kind of synthesis human face portrait quality evaluating method of feature based similarity is proposed, to improve the synthesis evaluating objective quality of human face portrait and the consistance of subjective quality assessment.

The technical scheme realizing the object of the invention comprises the steps:

(1) choose a width synthesis human face portrait to draw a portrait as test, choose original portrait corresponding to synthesis portrait as reference portrait; Test portrait being carried out pre-service with reference to drawing a portrait, making its size the same, pretreated test portrait and reference portrait are expressed as T (x), R (x);

(2) R (x) that test portrait T (x) and reference drawn a portrait is cut to non-overlapping portrait block in the same size, is expressed as TP _i(x) and RP _i(x), i=1,2 ..., N, N represent the number of portrait block, and x represents the position of pixel;

(3) quality assessment is carried out to the portrait block after segmentation:

(3a) Canny operator is utilized to calculate i-th test portrait block TP respectively _ithe marginal information E of (x) _t(x) and i-th reference portrait block RP _ithe marginal information E of (x) _r(x), and calculate E _t(x) and E _rthe similarity S of (x) _e(x):

$S_E\left(x\right)=\frac{2E_T\left(x\right).E_R\left(x\right)+T_1}{{E_T}^2\left(x\right)+{E_R}^2\left(x\right)+T_1},$

Wherein T ₁for being greater than the constant of 0;

(3b) i-th test portrait block TP is calculated respectively _ithe gradient magnitude G of (x) _t(x) and i-th reference portrait block RP _ithe gradient magnitude G of (x) _r(x), and calculate G _t(x) and G _rthe similarity S of (x) _g(x):

$S_G\left(x\right)=\frac{2G_T\left(x\right).G_R\left(x\right)+T_2}{{G_T}^2\left(x\right)+{G_R}^2\left(x\right)+T_2},$

Wherein T ₂for being greater than the constant of 0;

(3c) by similarity S that step (3a) obtains _ex similarity S that () and step (3b) obtain _gx () combines, obtain i-th test portrait block TP _ix () draws a portrait block RP with i reference _ithe similarity S of (x) _l(x):

S _L(x)＝(S _E(x)) ^α·(S _G(x)) ^β，

Wherein α is S _ethe factor of influence of (x), α >0; β is S _gthe factor of influence of (x), β >0;

(3d) according to the marginal information E that step (3a) obtains _t(x), E _rx similarity S that () and step (3c) obtain _lx () calculates portrait block quality assessment mark score:

$\mathrm{score}=\frac{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{S}_L\left(x\right)E_m\left(x\right)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{E}_m\left(x\right)},$

Wherein Ω is the spatial domain of whole portrait block, E _m(x)=max (E _t(x), E _r(x));

(4) quality assessment is carried out to view picture test portrait:

(4a) repeat step (3), until process N number of test portrait block, the quality assessment mark that each portrait block obtains is expressed as score _i, i=1,2 ..., N;

(4b) the quality assessment mark SCORE asking view picture to draw a portrait:

$\mathrm{SCORE}=\frac{1}{N}({\mathrm{score}}_1+{\mathrm{score}}_2+...+{\mathrm{score}}_i+...+{\mathrm{score}}_N).$

The present invention is owing to having taken into full account the singularity of synthesis human face portrait compared to natural image, achieve the evaluating objective quality of synthesis human face portrait, compared with traditional images quality evaluating method, the consistance that the method obtains synthesizing human face portrait quality evaluation result and subjective quality assessment is higher.

Simulation result shows: the present invention can extract edge and the Gradient Features of synthesis human face portrait preferably, in the quality assessment of synthesis human face portrait, obtain good effect, is a kind of quality evaluating method that effectively can be applicable to synthesize human face portrait.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention.

Embodiment

Core concept of the present invention is: for existing image quality evaluation algorithm application when synthesizing human face portrait quality assessment, evaluation result and the problem such as subjective quality assessment consistance is low, the human face portrait quality evaluating method of feature based similarity is proposed, to improve the subjective and objective consistance of image quality evaluation.Below provide the application example of this invention at synthesis human face portrait.

With reference to Fig. 1, the implementation step of this example is as follows:

Step 1, portrait pre-service.

From synthesis human face portrait storehouse, choose a width synthesis human face portrait as test portrait, draw a portrait corresponding reference draw a portrait from reference to choosing a width portrait storehouse with testing; Test portrait being carried out pre-service with reference to drawing a portrait, making its size identical, being namely 163 × 200 pixels, pretreated test portrait and reference portrait are expressed as T (x), R (x).

Step 2, to synthesis portrait and portrait piecemeal respectively.

To be that the zero lap of L × L draws a portrait block for T (x) and R (x) are cut to size, be expressed as TP _i(x), RP _i(x), i=1,2 ..., N, N represent the number of portrait block, and x represents the position of pixel, L=60.

Step 3, the quality assessment of synthesis human face portrait block.

3.1) portrait Block edge patterns similarity is calculated;

Canny operator extraction i-th 3.1a) is utilized to test portrait block TP _ithe edge feature E of (x) _t(x) with i-th with reference to drawing a portrait block RP _ithe edge feature E of (x) _r(x):

E _T(x)＝edge(TP _i(x),'canny')

E _R(x)＝edge(RP _i(x),'canny')；

Wherein edge () function is the edge indicator function that matlab software carries, ' canny' represents and select the canny Operator Method in edge () function to carry out rim detection.

3.1b) calculate E _t(x) and E _rthe similarity S of (x) _e(x):

$S_E\left(x\right)=\frac{2E_T\left(x\right).E_R\left(x\right)+T_1}{{E_T}^2\left(x\right)+{E_R}^2\left(x\right)+T_1},$

Wherein T ₁=0.85;

3.2) portrait block amplitude Gradient Features similarity is calculated;

3.2a) calculate extraction i-th test portrait block TP _ithe gradient magnitude G of (x) _t(x) with i-th with reference to drawing a portrait block RP _ithe gradient magnitude G of (x) _r(x)

$G_T\left(x\right)=\sqrt{{G_{\mathrm{Tx}}}^2+{G_{\mathrm{Ty}}}^2}$

$G_R\left(x\right)=\sqrt{{G_{\mathrm{Rx}}}^2+{G_{\mathrm{Ry}}}^2},$

Wherein, G _tx=conv2 (TP _i(x), A, ' same'), G _ty=conv2 (TP _i(x), B, ' same'),

G _Rx＝conv2(RP _i(x),A,'same')，G _Ry＝conv2(RP _i(x),B,'same')，

A＝[3 0-3；10 0-10；3 0-3]/16，B＝[3 10 3；0 0 0；-3-10-3]/16，

The convolution function that conv2 () carries for matlab software, ' same' represents the G utilizing conv2 () function to obtain _txand G _tysize and TP _ix () is consistent, G _rxand G _rysize and RP _ix () is consistent;

3.2b) calculate G _t(x) and G _rthe similarity S of (x) _g(x):

$S_G\left(x\right)=\frac{2G_T\left(x\right).G_R\left(x\right)+T_2}{{G_T}^2\left(x\right)+{G_R}^2\left(x\right)+T_2}$

Wherein T ₂=160;

3.3) to step 3.1) S that obtains _e(x) and step 3.2) S that obtains _gx () combines, obtain i-th test portrait block TP _ix () draws a portrait block RP with i reference _ithe similarity S of (x) _l(x):

S _L(x)＝(S _E(x)) ^α·(S _G(x)) ^β，

Wherein α is S _ethe factor of influence of (x), α >0; β is S _gthe factor of influence of (x), β >0; Get α=β=1;

3.4) according to step 3.1) the marginal information E that obtains _t(x), E _r(x) and step 3.3) the similarity S that obtains _lx () calculates portrait block quality assessment mark score:

$\mathrm{score}=\frac{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{S}_L\left(x\right)E_m\left(x\right)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{E}_m\left(x\right)},$

Wherein Ω is the spatial domain of whole portrait block, E _m(x)=max (E _t(x), E _r(x)).

Step 4, calculates the quality assessment mark of test synthesis human face portrait.

4.1) repeat step 3, until process all N number of test portrait blocks, the portrait block quality assessment mark that it obtains is expressed as score _i, i=1,2 ..., N;

4.2) the quality assessment mark SCORE asking view picture to draw a portrait:

$\mathrm{SCORE}=\frac{1}{N}({\mathrm{score}}_1+{\mathrm{score}}_2+...+{\mathrm{score}}_i+...+{\mathrm{score}}_N).$

Effect of the present invention can be further illustrated by following experiment:

The inventive method is tested on existing available synthesis human face portrait storehouse, and contrasts with classic map image quality evaluation method FSIM, SSIM, GMSD, IW-SSIM method, feasibility of the present invention is described.

1, experiment condition and description of test

Realize the MATLAB 2010a that software environment of the present invention is the exploitation of Mathworks company of the U.S., computing machine used is the personal computer of 2G Hz.

The reference representation data storehouse used in experiment is CUHK Student, 223, Purdue AR representation data storehouse human face portrait, and label is 1-223.

Synthesis portrait in the synthesis human face portrait database used in experiment comprises following 5 kinds:

One is utilize to synthesize human face portrait algorithm process representation data storehouse every width based on LLE and draw a portrait the synthesis human face portrait obtained, and label is 1_1-223_1;

Two is utilize to synthesize every width on human face portrait algorithm process representation data storehouse based on MWF and draw a portrait the synthesis human face portrait obtained, and label is 1_2-223_2;

Three is utilize SNS to synthesize every width on human face portrait algorithm process representation data storehouse draw a portrait the synthesis obtained and draw a portrait, and label is 1_3-223_3;

Four utilize SNS-SVR to synthesize every width portrait on human face portrait algorithm process representation data storehouse obtains, and label is 1_4-223_4;

Five utilize TFSPS to synthesize every width portrait on human face portrait algorithm process representation data storehouse obtains, and label is 1_5-223_5; Totally 1115 width human face segmentation portraits.

2. the performance index of picture quality objective evaluation

Adopt relatively more conventional Pearson related coefficient PLCC, Spearman coefficient of rank correlation SRCC and Kendall coefficient of rank correlation KRCC, analysis chart is as the consistance of method for objectively evaluating and subjective assessment, and PLCC, SRCC, KRCC show that more greatly the subjective evaluation consistance of picture quality is higher.

3, experiment content

Choosing original painting in experiment as totally 60 former facial images that database label is 1-60 is reference portrait, and choosing label in synthesis representation data storehouse is that totally 300 portraits of 1_1-60_1,1_2-60_2,1_3-60_3,1_4-60_4,1_5-60_5 are as testing synthesis human face portrait.

The inventive method is utilized to calculate the quality assessment mark of synthesis human face portrait, and calculate its PLCC, SRCC and KRCC parameter, judge the consistance that itself and subjective quality assessment are evaluated, and contrast with full reference image quality appraisement method FSIM, SSIM, GSMD, IW-SSIM of classics, result is as table 1.

The subjective and objective Conformance Assessment result of table 1 difference synthesis human face portrait evaluation method

As seen from Table 1, SROCC average of the present invention, PLCC average, KRCC average improve 0.7%, 7%, 2.4% than FSIM method respectively, improve 16.5%, 23%, 103% than SSIM method respectively, improve 12%, 19.7%, 12.3% respectively than IW-SSIM.In addition in maximal value and minimum value, all increase significantly or with additive method, there is comparability.

To sum up, the present invention can obtain good effect in the quality assessment of synthesis human face portrait; The present invention with subjective consistency in more existing classic algorithm FSIM, SSIM, GSMD, IW-SSIM be all significantly improved, be a kind of method that effectively can be applicable to synthesize human face portrait quality assessment.

Claims (5)

1. a synthesis human face portrait quality evaluating method for feature based similarity, comprises the steps:

(1) choose a width synthesis human face portrait to draw a portrait as test, choose original portrait corresponding to synthesis portrait as reference portrait; Test portrait being carried out pre-service with reference to drawing a portrait, making its size the same, pretreated test portrait and reference portrait are expressed as T (x), R (x);

(2) R (x) that test portrait T (x) and reference drawn a portrait is cut to non-overlapping portrait block in the same size, is expressed as TP _i(x) and RP _i(x), i=1,2 ..., N, N represent the number of portrait block, and x represents the position of pixel;

(3) quality assessment is carried out to the portrait block after segmentation:

(3a) Canny operator is utilized to calculate i-th test portrait block TP respectively _ithe marginal information E of (x) _t(x) and i-th reference portrait block RP _ithe marginal information E of (x) _r(x), and calculate E _t(x) and E _rthe similarity S of (x) _e(x):

$S_E\left(x\right)=\frac{2E_T\left(x\right).E_R\left(x\right)+T_1}{{E_T}^2\left(x\right)+{E_R}^2\left(x\right)+T_1},$

Wherein T ₁for being greater than the constant of 0;

(3b) i-th test portrait block TP is calculated respectively _ithe gradient magnitude G of (x) _t(x) and i-th reference portrait block RP _ithe gradient magnitude G of (x) _r(x), and calculate G _t(x) and G _rthe similarity S of (x) _g(x):

$S_G\left(x\right)=\frac{2G_T\left(x\right).G_R\left(x\right)+T_2}{{G_T}^2\left(x\right)+{G_R}^2\left(x\right)+T_2},$

Wherein T ₂for being greater than the constant of 0;

(3c) by similarity S that step (3a) obtains _ex similarity S that () and step (3b) obtain _gx () combines, obtain i-th test portrait block TP _ix () draws a portrait block RP with i reference _ithe similarity S of (x) _l(x):

S _L(x)＝(S _E(x)) ^α·(S _G(x)) ^β，

Wherein α is S _ethe factor of influence of (x), α >0; β is S _gthe factor of influence of (x), β >0;

(3d) according to the marginal information E that step (3a) obtains _t(x), E _rx similarity S that () and step (3c) obtain _l(x)

Calculate portrait block quality assessment mark score:

$\mathrm{score}=\frac{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{S}_L\left(x\right)E_m\left(x\right)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{E}_m\left(x\right)},$

Wherein Ω is the spatial domain of whole portrait block, E _m(x)=max (E _t(x), E _r(x));

(4) quality assessment is carried out to view picture test portrait:

(4a) repeat step (3), until process N number of test portrait block, the quality assessment mark that each portrait block obtains is expressed as score _i, i=1,2 ..., N;

(4b) the quality assessment mark SCORE asking view picture to draw a portrait:

$\mathrm{SCORE}=\frac{1}{N}({\mathrm{score}}_1+{\mathrm{score}}_2+...+{\mathrm{score}}_i+...+{\mathrm{score}}_N).$

2. the synthesis human face portrait quality evaluating method of feature based similarity according to claim 1, is characterized in that, calculates i-th test portrait block TP in step (3a) _ithe marginal information E of (x) _t(x), by following formulae discovery:

E _T(x)＝edge(TP _i(x),'canny')；

Wherein edge () function is the edge indicator function that matlab software carries, ' canny' represents and select the canny Operator Method in edge () function to carry out rim detection.

3. the synthesis human face portrait quality evaluating method of feature based similarity according to claim 1, is characterized in that, calculates i-th with reference to portrait block RP in step (3a) _ithe marginal information E of (x) _r(x), by following formulae discovery:

E _R(x)＝edge(RP _i(x),'canny')；

Wherein edge () function is the edge indicator function that matlab software carries, ' canny' represents and select the canny Operator Method in edge () function to carry out rim detection.

4. the synthesis human face portrait quality evaluating method of feature based similarity according to claim 1, is characterized in that, calculates i-th test portrait block TP in step (3b) _ithe gradient magnitude G of (x) _t(x), by following formulae discovery:

$G_T\left(x\right)=\sqrt{{G_{\mathrm{Tx}}}^2+{G_{\mathrm{Ty}}}^2};$

Wherein G _tx=conv2 (TP _i(x), A, ' same'), G _ty=conv2 (TP _i(x), B, ' same'),

A=[3 0-3; 10 0-10; 3 0-3]/16, B=[3 10 3; 000;-3-10-3]/16, conv2 () convolution function of carrying for matlab software, ' same' represents the G utilizing conv2 () function to obtain _txand G _tysize and TP _ix () is consistent.

5. the synthesis human face portrait quality evaluating method of feature based similarity according to claim 1, is characterized in that, calculates i-th with reference to portrait block RP in step (3b) _ithe gradient magnitude G of (x) _r(x), by following formulae discovery:

$G_R\left(x\right)=\sqrt{{G_{\mathrm{Rx}}}^2+{G_{\mathrm{Ry}}}^2};$

Wherein G _rx=conv2 (RP _i(x), A, ' same'), G _ry=conv2 (RP _i(x), B, ' same')

A=[3 0-3; 10 0-10; 3 0-3]/16, B=[3 10 3; 000;-3-10-3]/16, conv2 () convolution function of carrying for matlab software, ' same' represents the G utilizing conv2 () function to obtain _rxand G _rysize and RP _ix () is consistent.