CN101430760A - Human face super-resolution processing method based on linear and Bayesian probability mixed model - Google Patents

Abstract

The invention relates to a face super-resolution processing method based on a linear and Bayesian probability mixed model, which consists of image acquisition and image preprocessing, face image principal component extraction, eigenface super-resolution processing, principal component super-resolution processing and Bayesian probability correction. The face super-resolution processing method adopts a mixed model based on a linear model and a Bayesian probability model, firstly, main component components of a low-resolution face image are extracted, and a super-resolution result of a low-resolution eigenface is obtained through training by utilizing the linear model to realize a primary super-resolution result of the main components; and then, a Bayes probability model is utilized, and the preliminary super-resolution result is corrected through maximum posterior probability estimation, so that the robustness of the super-resolution algorithm to noise is effectively enhanced, and the definition of the super-resolution result is improved.

Description

Human face super-resolution processing method based on linear and Bayesian probability mixed model

Technical field

The present invention relates to human face super-resolution processing method, belong to intelligent monitoring technology in computer vision, particularly face recognition technology based on linear and Bayesian probability mixed model.

Background technology

People's face is one of human most important biological characteristic, has reflected a lot of important biological information, as identity, sex, age, race, expression or the like.Along with fast development of computer technology, the hot issue that also becomes recent researches based on the computer vision and the pattern recognition problem of facial image.Comprising all kinds of identification problems such as human face region detection, recognition of face, human face expression identifications.But, be subject to equipment cost, weather, and incident people face between shooting distance, captured facial image is often owing to resolution is crossed the low identification that is difficult to.And along with the reduction of facial image resolution, the discrimination of general face recognition algorithms will descend rapidly.And in many practical applications, often need again taken low resolution facial image, discern as facial image taken in the monitoring video.In order to improve the recognition performance to low resolution people face, general way is to utilize various conventional interpolator arithmetics, identification again after the single width facial image is simply amplified.Yet conventional interpolator arithmetic can not reconstruct the detailed information of real embodiment face characteristic, thereby there is no obvious help to improving discrimination.The Image Super Resolution Processing technology is meant that the face image super-resolution treatment technology is that the Image Super Resolution Processing technology is in people's application on the face by low-resolution image prediction high-definition picture.

The Image Super Resolution Processing technology mainly contains two big classes at present: based on the method for rebuilding with based on the method for learning.

1 based on the SUPERRESOLUTION PROCESSING FOR ACOUSTIC technology of rebuilding

Mainly be divided into two class methods again based on the SUPERRESOLUTION PROCESSING FOR ACOUSTIC technology of rebuilding: frequency domain method and spatial domain method.Frequency domain method is actually and solves the image interpolation problem in frequency domain, and it observes the shift characteristics that model is based on Fourier transform.In the class methods of spatial domain, its linear spatial domain observation model relates to the overall situation and local motion, optical dimming, intraframe motion blur, spatially-variable point spread function, imperfect sampling, compression artefacts and some other content.

2 SUPERRESOLUTION PROCESSING FOR ACOUSTIC technology based on study

Method based on study is a kind of method based on the identification priori.Go study identification to specify classification by algorithm, as object, scene, image, the identification priori that obtains is used for super-resolution, has obtained the better result of the super-resolution method based on reconstruct than the level and smooth priori of traditional use standard, Here it is based on the super-resolution method of learning.

But always there be the problem not high to the robustness of noise in the SUPERRESOLUTION PROCESSING FOR ACOUSTIC technology that is based on reconstruction, and always higher based on the complexity of the SUPERRESOLUTION PROCESSING FOR ACOUSTIC technology algorithm of learning.

The human face super-resolution treatment technology can be applied in a lot of fields, as recognition of face etc.Facial image in the video monitoring is smaller usually, and in order to discern them better, can carry out super-resolution to them earlier and amplify, and then discern, and the human face expression analysis.In some scene, as cameo shot, facial image is too little, to such an extent as to the analysis of can't expressing one's feelings can be carried out super-resolution to it earlier equally and amplify the analysis of expressing one's feelings again.Because people are familiar with people's face and are responsive very much, thereby the face image super-resolution scheduling algorithm is studied is a challenging problem.

Summary of the invention

The objective of the invention is to set up a cover face image super-resolution method, be used for the low resolution facial image is carried out superresolution restoration, and improve the robustness of human face super-resolution system noise.

In order to achieve the above object, technical solution of the present invention provides the method for a cover face image super-resolution, is made up of image acquisition and image pre-service, the extraction of facial image major component, eigenface SUPERRESOLUTION PROCESSING FOR ACOUSTIC, people's face major component SUPERRESOLUTION PROCESSING FOR ACOUSTIC and Bayesian probability correction.Image acquisition refers to the facial image absorption.The image pre-service refers to the yardstick normalization of image and adopts the gray scale normalization of gray balance and grey level stretching.The facial image major component is extracted extraction and the selection that comprises the facial image major component, main overall scatter matrix by the structure training image, pass through the eigenwert and the proper vector of the method calculated population scatter matrix of svd then, sort from big to small by eigenwert, select the eigenface of preceding L eigenwert characteristic of correspondence vector as major component, have 100 eigenwerts after extraction is handled through major component for the low resolution facial image, L=30 was as the number of selected eigenwert before the present invention adopted.The eigenface SUPERRESOLUTION PROCESSING FOR ACOUSTIC is the major component of the image in the people's face training storehouse that will select, and promptly eigenface is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC, mainly comprises the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result who uses linear model calculated characteristics face.People's face major component SUPERRESOLUTION PROCESSING FOR ACOUSTIC refers to that mainly the major component of the low-resolution image that will gather carries out SUPERRESOLUTION PROCESSING FOR ACOUSTIC.The Bayesian probability correction is meant the non-major component part with facial image in the Bayesian Estimation device correction eigenface ultra-resolution method of a maximum a posteriori probability, and the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of major component and non-major component is sued for peace obtains the final SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of low resolution facial image then.

The present invention is based on the human face super-resolution method of linear and Bayesian probability mixed model, principal feature of the present invention is when utilizing linear model that the main composition of gathering the low resolution facial image is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC, utilize Bayesian model that the non-major component of the low resolution facial image of collection is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC, and then reach the purpose of facial image super-resolution.The main performing step of the method for the invention is as follows:

Step 1: image acquisition and image pre-service;

Use image capture device to gather the low resolution facial image, and reading high-resolution human face image and low resolution facial image in high-resolution human face image training storehouse and the low resolution facial image training storehouse respectively, the method for using histogram equalization and histogram to stretch is done pre-service to the above-mentioned three-type-person's face image that collects respectively;

Step 2: the facial image major component is extracted;

The pre-service result who obtains training low resolution facial image in the storehouse from step 1 is carried out the major component of facial image and extract, promptly extract low resolution facial image eigenface;

Step 3: the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of facial image eigenface;

Mainly realize, in linear model, use the high-resolution human face image in the training set to substitute the low resolution facial image, realize SUPERRESOLUTION PROCESSING FOR ACOUSTIC low resolution facial image eigenface by setting up linear model;

Step 4: facial image major component SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

Utilize the result of the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of the resulting low resolution facial image of step 3 eigenface, the major component of the low resolution facial image gathered is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

Step 5: Bayesian probability correction;

Use the Bayesian Estimation device of a maximum a posteriori probability that the non-major component of facial image in the step four-player face major component super-resolution is partly revised, obtain the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of final facial image.

By the associating of above five steps, can carry out SUPERRESOLUTION PROCESSING FOR ACOUSTIC to the low resolution facial image.

The invention has the advantages that:

(1) the present invention adopts the human face super-resolution method based on linear and Bayesian probability mixed model, except adopting linear model, has also adopted the Bayesian probability correction model;

(2) the resulting result of mixture model is far superior to the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result that only obtains by linear model among the present invention, and is very high to the robustness of noise, can recover original high-resolution human face image under the noise energy condition with higher;

(3) the resulting result of mixture model is far superior to the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result that only obtains by the Bayesian probability model among the present invention, and the high-resolution human face image and the original high resolution facial image of recovery are closely similar.

Description of drawings

Fig. 1 is the synoptic diagram of face image super-resolution method of the present invention;

Fig. 2 is the low eigenface of differentiating facial image of the present invention;

Fig. 3 is the eigenface and the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result images corresponding with it of low resolution facial image of the present invention;

Fig. 4 is inventor's face Image Super Resolution Processing result.

Embodiment

Below in conjunction with accompanying drawing the present invention is described in further details.

Figure 1 shows that the synoptic diagram of the human face super-resolution method that the present invention is based on linear and Bayesian probability mixed model, the method for the invention comprises following steps:

Step 1: image acquisition and image pre-service;

Use image capture device to gather the low resolution facial image, and reading high-resolution human face image and low resolution facial image in high-resolution human face image training storehouse and the low resolution facial image training storehouse respectively, the method for using histogram equalization and histogram to stretch is done pre-service to the above-mentioned three-type-person's face image that collects respectively;

Step 2: the facial image major component is extracted;

The pre-service result who obtains training low resolution facial image in the storehouse from step 1 is carried out the major component of facial image and extract, promptly extract low resolution facial image eigenface;

At first, will from step 1, obtain training the pre-service result of low resolution facial image in the storehouse to change into vector, construct overall scatter matrix;

Then, by the method for svd, the eigenwert of calculated population scatter matrix and proper vector;

At last, eigenwert according to ordering from big to small, is selected the major component of preceding L eigenwert characteristic of correspondence vector as people's face, i.e. eigenface;

Have 100 eigenwerts after extraction is handled through major component for the low resolution facial image, L=30 was as the number of selected eigenwert before the present invention adopted;

Step 3: the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of facial image eigenface;

Mainly realize, in linear model, use the high-resolution human face image in the training set to substitute the low resolution facial image, realize SUPERRESOLUTION PROCESSING FOR ACOUSTIC low resolution facial image eigenface by setting up linear model;

Its concrete steps are as follows:

(1) method by svd, the eigenvectors matrix of the overall scatter matrix of low resolution facial image in the calculation training sample:

$E={\mathrm{UV\Λ}}^{-\frac{1}{2}}$

Wherein, E is an eigenvectors matrix, and U is a training of human face image array, and V is the eigenvectors matrix of the transposition of overall scatter matrix, and Λ is an eigenvalue matrix.

(2) eigenface of low resolution is expressed as the linearity summation of low resolution training image:

$e_i^l=U_l{V}_l{\mathrm{\Λ}}_l^{-\frac{1}{2}}{a}_i+m_l$

Wherein, Be low resolution eigenface, U _lBe low resolution training of human face image array, V _lBe the eigenvectors matrix of the transposition of the overall scatter matrix of low resolution training facial image, Λ _lBe the eigenvalue matrix of low resolution training facial image, a _iBe weighting coefficient, m _lBe low resolution average face vector, as shown in Figure 2.

(3) replace with the high-resolution human face image of training in the facial image database and the low resolution facial image of correspondence, can obtain the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the low resolution eigenface of correspondence, as shown in Figure 3:

$s_i^h=U_h{V}_l{\mathrm{\Λ}}_l^{-\frac{1}{2}}{a}_i+m_h$

Wherein,

Be the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of low resolution eigenface, U _hBe high resolving power training of human face image array, m _hBe the average face vector of high-resolution human face image, a _i, V _l, Λ _lBe respectively the parameter of low resolution facial image.

Step 4: facial image major component SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

Utilize the result of the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of the resulting facial image eigenface of step 3, the major component of the low resolution facial image gathered is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC, its main performing step is as follows:

(1) with the low resolution facial image projection of gathering on the major component of low resolution facial image, with this low resolution facial image x _lBe defined as two parts, wherein

Be defined as low resolution facial image major component linearity and,

Be defined as all the other non-major component parts of low resolution facial image:

$x_l=x_l^1+x_l^2$

$x_l^1=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{e}_i^l+m_l$

$x_l^2=e_{\mathrm{lx}}$

Wherein, a _iFor the facial image major component linear and in weighting coefficient, by obtaining as facial image major component projection process among Fig. 1, L is the number of selected weighting coefficient,

Be low resolution eigenface, e _LxAll the other non-major component parts for the low resolution facial image.

(2) the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of resulting low resolution eigenface calculated the preliminary SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the low resolution facial image of collection during the use characteristic face super-resolution was handled:

$x_h^1=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{s}_i^h+m_h$

Wherein,

For low resolution facial image major component linear and SUPERRESOLUTION PROCESSING FOR ACOUSTIC result, SUPERRESOLUTION PROCESSING FOR ACOUSTIC result for low resolution eigenface correspondence.

Step 5: Bayesian probability correction;

The Bayesian Estimation device of a maximum a posteriori probability of use is estimated SUPERRESOLUTION PROCESSING FOR ACOUSTIC results of all the other non-major component parts of low resolution facial image in the step 4, preliminary SUPERRESOLUTION PROCESSING FOR ACOUSTIC result to the low resolution facial image that obtains in the step 4 revises with this estimated value, SUPERRESOLUTION PROCESSING FOR ACOUSTIC result's summation with major component and non-major component, obtain final face image super-resolution result, concrete steps are finished as follows:

Non-major component partly is:

$x_l^2=e_{\mathrm{lx}}$

It satisfies equation:

$x_l^2=e_{\mathrm{lx}}={\mathrm{Hx}}_h^2+n={\mathrm{HE}}_h{a}_h{+\mathrm{He}}_{\mathrm{hx}}+n$

Wherein, e _HxBe e _LxSUPERRESOLUTION PROCESSING FOR ACOUSTIC result's non-major component part,

Be the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the non-major component part of low resolution facial image, E _hThe eigenvectors matrix of expression high-resolution human face image, a _hThe weighting coefficient of expression high-resolution human face image, H is the image degradation matrix, n is a noise, further can obtain it through conversion and satisfy equation:

$E_l^T{\mathrm{HE}}_h{a}_h+E_l^{T}H{e}_{\mathrm{hx}}+E_l^{T}n=0$

Wherein,

The transposition of the eigenvectors matrix of expression low resolution facial image, the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result who calculates the non-major component part of facial image only need solve above equation, the Bayesian Estimation device of the maximum a posteriori probability that the present invention uses:

${\hat{a}}_h=\underset{a_h}{\arg \max }\left\{p\left({\hat{a}}_h\right|a_h)p\left(a_h\right)\right\}$

Wherein,

The estimated value of expression high-resolution human face image weighting coefficient, p (a _h) be prior probability,

It is conditional probability.

The prior probability of facial image can be regarded a Gauss model as:

$p\left(a_h\right)=\frac{1}{Z}\exp (-{(a_h-{\mathrm{\μ}}_a)}^T{\mathrm{\Λ}}^{-1}(a_h-{\mathrm{\μ}}_a))$

Wherein, Z is a normalization coefficient, and Λ is high-resolution human face image weighting coefficient a _hCovariance matrix, μ _aBe high-resolution human face image weighting coefficient a _hAverage.Simultaneously with F=He _Hx+ n regards another Gauss model as, and then the prior probability of F is:

$p\left(F\right)=\frac{1}{Z}\exp (-{(F-{\mathrm{\μ}}_F)}^T{K}^{-1}(F-{\mathrm{\μ}}_F))$

Wherein, Z is a normalization coefficient, and K is the covariance matrix of F, μ _FIt is the average of F.K and μ _FCan calculate by training sample, can obtain high-resolution human face image weighting coefficient at last

Maximum a posteriori probability be estimated as:

${\hat{a}}_h=\underset{a_h}{\arg \min }\{{({\hat{a}}_h-E_l^T{\mathrm{HE}}_h{a}_h-E_i^T{\mathrm{\μ}}_F)}^T$

$Q^{-1}({\hat{a}}_h-E_l^T{\mathrm{HE}}_h{a}_h-E_l^T{\mathrm{\μ}}_F)$

$+(a_h-{\mathrm{\μ}}_a){\mathrm{\Λ}}^{-1}(a_h-{\mathrm{\μ}}_a)\}$

Wherein, Q is the covariance matrix that the transposed matrix of E is multiplied by the matrix that F obtains, and the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result who then gathers the non-major component part of low resolution facial image is:

$x_h^2=E_h{\hat{a}}_h$

E _hCan obtain by the training of the high-resolution human face image in the training image storehouse, and in the super-resolution of non-major component part, the error that the subspace is represented can be ignored, so $x_h^2=E_h{\hat{a}}_h$ Can regard exact solution as, with the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of major component and non-major component summation, the final SUPERRESOLUTION PROCESSING FOR ACOUSTIC of then gathering low-resolution image is x as a result _hAs follows:

$x_h=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{s}_i^h+m_h+E_h{\hat{a}}_h$

Training face database of the present invention adopts classical U.S.'s face database (Facial RecognitionTechnology, abbreviate FERET as) face database, picked at random 126 256 gray level images, be normalized to 40 * 40 image as the high-resolution human face image in the training sample database, then by down-sampled 10 * 10 the low resolution facial image that obtains, add Gaussian noise (m=0 simultaneously, σ=0.1, wherein, m is the noise average, and σ is the noise mean square deviation) obtain in the training sample database corresponding low resolution facial image.Have 100 eigenwerts after extraction is handled through major component for the low resolution facial image, L=30 was as the number of selected eigenwert before the present invention adopted, and concrete experimental result as shown in Figure 4.

Among Fig. 4, (a) be the low resolution facial image of input, (b) be the high-resolution human face image that obtains by bicubic interpolation, (c) be the high-resolution human face image that only obtains by linear model, (d) be the resulting super-resolution facial image of mixture model by linear and Bayesian Estimation of the present invention, (e) be original high-resolution human face image, come as can be seen from the results, the resulting result of mixture model is far superior to the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result that only obtains by linear model, and closely similar with the original high resolution facial image.

Simultaneously, it is as shown in table 1 to contrast this several method gained result's signal to noise ratio (S/N ratio):

Table 1

Signal to noise ratio (S/N ratio)	Bicubic interpolation	Linear model	The present invention
				Image 1	12.5944	15.6025	17.1287
Image 2	12.4711	17.1703	18.0258
				Image 3	12.5078	17.4361	17.5645
Image 4	13.0851	18.0217	19.8149

As seen from Table 1, human face super-resolution method based on linear and Bayesian probability mixed model used in the present invention all is better than other two kinds of methods to the signal to noise ratio (S/N ratio) result of the low resolution facial image of four kinds of inputs, the present invention is higher to the robustness of noise, can recover original high-resolution facial image under the noise energy condition with higher.

Claims (2)

1, based on the human face super-resolution processing method of linearity and Bayesian probability mixed model, this method comprises:

Step 1: image acquisition and image pre-service;

Use image capture device to gather the low resolution facial image, and reading high-resolution human face image and low resolution facial image in high-resolution human face image training storehouse and the low resolution facial image training storehouse respectively, the method for using histogram equalization and histogram to stretch is done pre-service to the above-mentioned three-type-person's face image that collects respectively;

Step 2: the facial image major component is extracted;

The pre-service result who obtains training low resolution facial image in the storehouse from step 1 is carried out the major component of facial image and extract, promptly extract low resolution facial image eigenface;

At first, will from step 1, obtain training the pre-service result of low resolution facial image in the storehouse to change into vector, construct overall scatter matrix;

Then, by the method for svd, the eigenwert of calculated population scatter matrix and proper vector;

At last, eigenwert according to ordering from big to small, is selected the major component of preceding L eigenwert characteristic of correspondence vector as people's face, i.e. eigenface;

Step 3: the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of facial image eigenface;

Set up linear model, in linear model, use the high-resolution human face image in the training set to substitute the low resolution facial image, promptly low resolution facial image eigenface is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

Its concrete steps are as follows:

(1) method by svd, the eigenvectors matrix of the overall scatter matrix of low resolution facial image in the calculation training sample:

$E=\mathrm{UV}{\mathrm{\Λ}}^{-\frac{1}{2}}$

Wherein, E is an eigenvectors matrix, and U is a training of human face image array, and V is the eigenvectors matrix of the transposition of overall scatter matrix, and Λ is an eigenvalue matrix;

(2) eigenface of low resolution is expressed as the linearity summation of low resolution training image:

$e_i^l=U_l{V}_l{\mathrm{\Λ}}_l^{-\frac{1}{2}}{a}_i+m_l$

Wherein,

Be low resolution eigenface, U _lBe low resolution training of human face image array, V _lBe the eigenvectors matrix of the transposition of the overall scatter matrix of low resolution training facial image, Λ _lBe the eigenvalue matrix of low resolution training facial image, a _iBe weighting coefficient, m _lBe low resolution average face vector;

(3) replace with the high-resolution human face image of training in the facial image database and the low resolution facial image of correspondence, obtain the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the low resolution eigenface of correspondence:

$s_i^h=U_h{V}_l{\mathrm{\Λ}}_l^{-\frac{1}{2}}{a}_i+m_h$

Wherein,

Be the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of low resolution eigenface, U _hBe high resolving power training of human face image array, m _hBe the average face vector of high-resolution human face image, a _i, V _l, Λ _lBe respectively the parameter of low resolution facial image;

Step 4: facial image major component SUPERRESOLUTION PROCESSING FOR ACOUSTIC;

Utilize the result of the SUPERRESOLUTION PROCESSING FOR ACOUSTIC of the resulting low resolution eigenface of step 3, the major component of the low resolution facial image gathered is carried out SUPERRESOLUTION PROCESSING FOR ACOUSTIC, its step is as follows:

(1) with the low resolution facial image projection of gathering on the major component of low resolution facial image, with this low resolution facial image x _lBe defined as two parts, wherein

Be defined as low resolution facial image major component linearity and,

Be defined as all the other non-major component parts of low resolution facial image:

$x_l=x_l^1+x_l^2$

$x_l^1=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{e}_i^l+m_l$

$x_l^2=e_{\mathrm{lx}}$

Wherein, a _iFor the facial image major component linear and in weighting coefficient, obtain by the projection of facial image major component, L is the number of selected weighting coefficient,

Be low resolution eigenface, e _LxAll the other non-major component parts for the low resolution facial image;

(2) the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of resulting low resolution eigenface calculated the preliminary SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the low resolution facial image of collection during the use characteristic face super-resolution was handled:

$x_h^1=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{s}_i^h+m_h$

Wherein,

For low resolution facial image major component linear and SUPERRESOLUTION PROCESSING FOR ACOUSTIC result,

SUPERRESOLUTION PROCESSING FOR ACOUSTIC result for low resolution eigenface correspondence;

Step 5: Bayesian probability correction;

The Bayesian Estimation device of a maximum a posteriori probability of use is estimated SUPERRESOLUTION PROCESSING FOR ACOUSTIC results of all the other non-major component parts of low resolution facial image in the step 4, preliminary SUPERRESOLUTION PROCESSING FOR ACOUSTIC result to the low resolution facial image that obtains in the step 4 revises with this estimated value, SUPERRESOLUTION PROCESSING FOR ACOUSTIC result's summation with major component and non-major component, obtain final face image super-resolution result, concrete steps are as follows:

Non-major component partly is:

$x_l^2=e_{\mathrm{lx}}$

It satisfies equation:

$x_l^2=e_{\mathrm{lx}}={\mathrm{Hx}}_h^2+n={\mathrm{HE}}_h{a}_h+{\mathrm{He}}_{\mathrm{hx}}+n$

Wherein, e _HxBe e _LxSUPERRESOLUTION PROCESSING FOR ACOUSTIC result's non-major component part, Be the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of the non-major component part of low resolution facial image, E _hThe eigenvectors matrix of expression high-resolution human face image, a _hThe weighting coefficient of expression high-resolution human face image, H is the image degradation matrix, n is a noise, further obtains it through conversion and satisfies equation:

$E_l^T{\mathrm{HE}}_h{a}_h+E_l^T{\mathrm{He}}_{\mathrm{hx}}+E_l^{T}n=0$

Wherein, The transposition of the eigenvectors matrix of expression low resolution facial image;

The Bayesian Estimation device of the maximum a posteriori probability that the present invention uses:

${\hat{a}}_h=\underset{a_h}{\arg \max }\left\{p\left({\hat{a}}_h\right|a_h)p\left(a_h\right)\right\}$

Wherein,

The estimated value of expression high-resolution human face image weighting coefficient, p (a _h) be prior probability,

It is conditional probability;

The prior probability of facial image is counted as a Gauss model:

$p\left(a_h\right)=\frac{1}{Z}\exp (-{(a_h-{\mathrm{\μ}}_a)}^T{\mathrm{\Λ}}^{-1}(a_h-{\mathrm{\μ}}_a))$

Wherein, Z is a normalization coefficient, and Λ is high-resolution human face image weighting coefficient a _hCovariance matrix, μ _aBe high-resolution human face image weighting coefficient a _hAverage;

With F=He _Hx+ n regards another Gauss model as, and then the prior probability of F is:

$p\left(F\right)=\frac{1}{Z}\exp (-{(F-{\mathrm{\μ}}_F)}^T{K}^{-1}(F-{\mathrm{\μ}}_F))$

Wherein, Z is a normalization coefficient, and K is the covariance matrix of F, μ _FIt is the average of F;

K and μ _FCalculate by training sample, obtain high-resolution human face image weighting coefficient at last Maximum a posteriori probability be estimated as:

${\hat{a}}_h=\underset{a_h}{\arg \min }\{{({\hat{a}}_h-E_l^T{\mathrm{HE}}_h{a}_h-E_l^T{\mathrm{\μ}}_F)}^T$

$Q^{-1}({\hat{a}}_h-E_l^T{\mathrm{HE}}_h{a}_h-E_l^T{\mathrm{\μ}}_F)$

$+(a_h-{\mathrm{\μ}}_a){\mathrm{\Λ}}^{-1}(a_h-{\mathrm{\μ}}_a)\}$

The SUPERRESOLUTION PROCESSING FOR ACOUSTIC result who then gathers the non-major component part of low resolution facial image is:

$x_h^2=E_h{\hat{a}}_h$

E _hObtain by the training of the high-resolution human face image in the training image storehouse, and in the super-resolution of non-major component part, the error that the subspace is represented can be left in the basket, so will $x_h^2=E_h{\hat{a}}_h$ As exact solution, with the SUPERRESOLUTION PROCESSING FOR ACOUSTIC result of major component and non-major component summation, the final SUPERRESOLUTION PROCESSING FOR ACOUSTIC of then gathering low-resolution image is x as a result _hFor:

$x_h=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{a}_i{s}_i^h+m_h+E_h{\hat{a}}_h.$

2, the human face super-resolution processing method based on linearity and Bayesian probability mixed model according to claim 1, it is characterized in that, preceding L the eigenwert characteristic of correspondence vector of selection described in the step 2 is as the major component of people's face, have 100 eigenwerts after extraction is handled through major component for the low resolution facial image, adopt preceding 30 numbers that are L=30 as selected eigenwert.

Images