CN101697197B - Method for recognizing human face based on typical correlation analysis spatial super-resolution - Google Patents

Abstract

The invention discloses a method for recognizing human face based on typical correlation analysis spatial super-resolution. Aiming at the problem of the low recognition efficiency of a low-resolution face image, the invention provides a method for obtaining the recognition feature of the low-resolution face image in a high-resolution space by using the super-resolution reconstruction of a recognition feature. Based on a manifold learning theory, the recognition features of the high-resolution and low-resolution face images are considered to be generated by a common internal structure, so the method comprises the following steps: enhancing the consistency of neighborhood relationships of the recognition features of the high-resolution and low-resolution face images by using typical correlation analysis so as to better meet the hypothesis on a neighborhood reconstruction concept; and utilizing neighborhood reconstruction to obtain the recognition feature of the high-resolution face image corresponding to the tested low-resolution face image in a related subspace obtained by the transformation of the typical correlation analysis, and finally utilizing the feature to recognize a face. Experiments show that a recognition rate obtained by the method is less influenced by the resolution of the face image and is relatively higher.

Description

A kind of face identification method based on typical correlation analysis spatial super-resolution

Technical field

The present invention relates to the recognition of face field, concrete relate to a kind of low resolution face identification method based on the canonical correlation analysis space.

Background technology

Face recognition technology has very important significance in the safe-guard system of current society.As one of the main direction of studying in pattern-recognition and machine learning field, there has been a large amount of face recognition algorithms to be suggested.But because the restriction of distance and hardware condition etc., the facial image resolution interested of in the large scene video monitoring system, taking is often lower, has reduced the performance of recognition of face.How under the low resolution condition, to improve recognition effect, be the problem that recognition of face aspect in recent years need solve.

(Super-resolution SR) is meant and utilizes certain algorithm (low resolution obtains a width of cloth or a series of high-definition pictures (highresolution, technology HR) in LR) from a width of cloth or a series of low-resolution images image super-resolution.Therefore, the face image super-resolution algorithm is used as one of solution that improves low resolution facial image recognition effect very naturally.Particularly, utilize the super-resolution rebuilding algorithm to obtain visual effect facial image preferably earlier, and then carry out recognition of face.But face image super-resolution rebuilding part independently is carried out with identification division in this scheme, and the target localization of super-resolution improving image visual effect, rather than the raising discrimination.

Based on above reason, people such as Gunturk have proposed the face identification method based on eigenface territory super-resolution rebuilding.This method no longer attempts to improve the visual effect of image, but directly rebuilds the required high-resolution features face domain information of face identification system.At first adopt the KL conversion to extract the facial image eigenface; Use imaging model to set up the contact between the high low-resolution image characteristic; Suppose the prior probability of the high-resolution features of asking vector be the associating Gaussian distribution; Utilize the iteration steepest descent algorithm that this proper vector is carried out maximum a posteriori probability and estimate that the proper vector that last basis is tried to achieve is carried out recognition of face.This method provides a kind of and has well directly utilized super-resolution to carry out the framework of recognition of face, but this method computation complexity is higher, and the method has the space of further improvement on performance.

People such as Pablo have proposed the low resolution face identification method that super-resolution rebuilding and feature extraction combine.The characteristic of utilizing classical face characteristic extraction algorithm acquisition training image earlier is as prior imformation; The sorter objective function that utilizes this prior imformation then and be used to discern is set up the regularization objective function; And obtain the parameter of regularization objective function through the method for study, in application, come to obtain simultaneously super-resolution rebuilding result and recognition result with regularization objective function minimum.The regularization objective function model that this method proposed clear expression simultaneously is to the restriction of super-resolution rebuilding result and recognition result, but the learning process more complicated of objective function parameter.In order to reduce algorithm complex, to improve discrimination, need to seek better method and solve low resolution recognition of face problem.

Summary of the invention

The objective of the invention is to overcome the shortcoming of above-mentioned prior art, proposed a kind of based on canonical correlation analysis (Canonical Correlation Analysis, CCA) face identification method of spatial super-resolution.

In order to achieve the above object, the technical scheme of the present invention's employing is:

1) at first; Utilize classical feature extracting method to extract high low resolution training of human face image recognition characteristic, utilize two groups of recognition features being extracted as training data, based on the requirement of CCA algorithm; Must shine upon base vector, recognition feature is transformed into CCA correlator space based on this mapping base vector;

2) secondly, in the correlator space, utilize neighborhood reconstruct to try to achieve the test low resolution facial image recognition feature recognition feature corresponding, realize the super-resolution rebuilding of recognition feature in the high resolving power territory;

3) last, utilize nearest neighbor classifier based on the L2 norm, carry out the arest neighbors Classification and Identification according to the recognition feature in the CCA space of trying to achieve, thereby obtain discrimination.

Recognition feature of the present invention is transformed into CCA correlator space and may further comprise the steps:

1) establishing high resolving power and the low resolution training of human face image collection removed after the average is respectively $I^H={\left\{I_i^H\right\}}_{i=1}^m=[I_1^H,I_2^H,...,I_m^H],$ $I^L={\left\{I_i^L\right\}}_{i=1}^m=[I_1^L,I_2^L,...,I_m^L],$ I wherein _i ^HBe i panel height image in different resolution, I _i ^LBe i width of cloth low-resolution image, m is the total number of samples of training image; Definition ${\left(B^H\right)}^T{I}_i^H=x_i^H,$ B wherein ^HBe that the high-definition picture characteristic of correspondence is extracted matrix, x _i ^HIt is representative's face image I _i ^HProper vector; The character representation that in like manner, can obtain low-resolution image does $x_i^L={\left(B^L\right)}^T{I}_i^L,$ B wherein ^LBe that low resolution facial image characteristic of correspondence is extracted matrix, x _i ^LRepresent low resolution facial image I _i ^LProper vector; Thus, obtain representing that the set of eigenvectors of corresponding high low resolution training image does $X^H={\left\{x_i^H\right\}}_{i=1}^m$ With $X^L={\left\{x_i^L\right\}}_{i=1}^m,$ X wherein _i ^HBe i panel height resolution facial image characteristic of correspondence vector, x _i ^LIt is the proper vector of i width of cloth low resolution facial image;

2) for set of eigenvectors X ^H, X ^LDeduct its average x respectively ^HAnd x ^L, obtain the data set of centralization

With

Definition

Be respectively With

Auto-covariance matrix,

With Be respectively

With

The cross covariance matrix, wherein E [] represents mathematical expectation, T represents the transposition computing; Calculate ${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="H2009102075622E00021.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=C_{11}^{-1}{C}_{12}{C}_{22}^{-1}{C}_{21},$ $R_2=C_{22}^{-1}{C}_{21}{C}_{11}^{-1}{C}_{12};$ R ₁And R ₂Proper vector be the base vector V that asks ^HAnd V ^L

3) utilize the base vector of asking, with set of eigenvectors X ^H, X ^LTransform to the CCA subspace, obtain the corresponding projection coefficient collection of high resolving power and low resolution training facial image $C^H={\left\{c_i^H\right\}}_{i=1}^m,$ $C^L={\left\{c_i^L\right\}}_{i=1}^m,$ That is:

With

Recognition feature super-resolution rebuilding step of the present invention is following:

1) the low resolution people's face test pattern I to importing _l, try to achieve its characteristic of correspondence vector x _l, further with x _lTransform to the CCA subspace, obtain c _l, that is: c _l=(V ^L) ^T(x _l-x ^L);

2) according to the thought of neighborhood reconstruct, at C ^LThe middle c that seeks _lK nearest neighbor { c _Li ^L} _I=1 ^K, c wherein _Li ^LBe c _lI neighbour, and ask the reconstruct weights $W^G={\left\{w_i^G\right\}}_{i=1}^K,$ W wherein _i ^GBe the corresponding weights coefficient of i neighbour, make objective function $\mathrm{\&epsiv;}=\left|\right|c_l-\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G{c}_{\mathrm{Li}}^L\left|\right|$ Reach minimum, and satisfy $\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G=1;$

3) with weights $W^G={\left\{w_i^G\right\}}_{i=1}^K$ Be applied to C ^HIn with { c _Li ^L} _I=1 ^KCorresponding { c _Li ^H} _I=1 ^K, promptly reconstruct I _lThe recognition feature of corresponding full resolution pricture is: $c_h=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G{c}_{\mathrm{Li}}^H.$

Said classical feature extracting method adopts principal component analysis (PCA) or linear judgment analysis.

The present invention regards the recognition feature of high low resolution facial image as the variable of two different dimensions; Utilize canonical correlation analysis to extract the correlator space of high low resolution facial image recognition feature, thereby strengthen the consistance of high low resolution facial image recognition feature topological structure.In the correlator space, the present invention utilizes neighborhood reconstruct to try to achieve the recognition feature of the corresponding high-resolution human face image area of test low resolution facial image, and utilizes the nearest neighbor classifier based on the L2 norm to carry out Classification and Identification, thereby obtains discrimination.

Description of drawings

Fig. 1 algorithm frame of the present invention;

A certain personage's five width of cloth facial expression images in Fig. 2 CAS-PEAL expression storehouse;

Fig. 3 utilizes CAS-PEAL expression picture library gained discrimination result;

Fig. 4 utilizes LDA to extract characteristic gained discrimination result;

Fig. 5 AR facial image database example, wherein (a) is 32 * 32 high resolving power training facial image, (b) is 8 * 8 low resolution training facial image, (c) is 8 * 8 low resolution test facial image;

Fig. 6 proper vector dimension size is to the influence of discrimination.

Embodiment

For making the object of the invention, technical scheme and advantage clearer,, the present invention is explained further details below in conjunction with accompanying drawing and instantiation.These instances are only illustrative, and are not limitation of the present invention.

The face image super-resolution identification problem can be described as: known two corresponding each other high low resolution facial image training set I ^HAnd I ^LPerhaps two mutual training set X of corresponding facial image recognition feature ^HAnd X ^L, import a width of cloth low resolution facial image I _l, ask the recognition feature c of its corresponding high-resolution human face image _h

The theory hypothesis people face space of manifold learning is a kind of embedding manifold structure, promptly is embedded in the low dimension stream shape of higher-dimension theorem in Euclid space.This higher-dimension structure that shows that people's face data set constitutes is hanged down dimension theorem in Euclid space topological homeomorphism with certain under local sense.Only resolution is different between corresponding high low resolution people's face data set, and just dimension is different.Therefore, can think that high low resolution people's face data set is to be generated by relevant immanent structure.Satisfying under the assumed condition of manifold learning; Can be according to local linear (the Locallylinear embedding that embeds in the manifold learning; LLE) principle of algorithm, the thought of employing neighborhood reconstruct is carried out face image super-resolution rebuilding; The neighborhood and the reconstruct weights that are about to the test data that low-resolution spatial tries to achieve are applied to high resolution space, thereby reconstruct the high resolution space data of test high-resolution data correspondence.

In recent years,, proposed some human face super-resolution algorithms, obtained good effect based on manifold learning based on above theory.Equally, if regard the recognition feature of high low resolution facial image as the height dimension data respectively, have reason also to think that these two data sets are to be generated by relevant immanent structure.Therefore, also can use based on the high-resolution recognition feature of the direct reconstruction of the thinking of manifold learning.

But in reality,, often can not obtain well satisfying about the hypothesis of high high-resolution data afflux shape local topology homeomorphism owing to reasons such as training picture library limited in number.CCA is the method that is used for analyzing two groups of data line sexual intercourse at first, its objective is that being respectively every group of data seeks one group of base vector, makes to reach maximum through the data dependence after these two groups of base vector conversion.The present invention with the recognition feature of high-resolution and low-resolution image as two groups of pending data; Set up the correlator space through CCA; Consistance with the neighborhood relationships that strengthens high low resolution facial image recognition feature; Satisfying hypothesis better based on neighborhood reconstruct thought in the manifold learning method, thus reconstruct high-resolution human face image recognition characteristic more accurately.

Algorithm frame of the present invention is as shown in Figure 1, mainly comprises three parts: facial image feature extraction, the super-resolution rebuilding of recognition feature and final identifying.

If high resolving power after the removal average and low resolution training of human face image collection are respectively $I^H={\left\{I_i^H\right\}}_{i=1}^m=[I_1^H,I_2^H,...,I_m^H],$ $I^L={\left\{I_i^L\right\}}_{i=1}^m=[I_1^L,I_2^L,...,I_m^L],$ Wherein m is a number of samples.The facial image characteristic can through classical principal component analysis (PCA) (Principal Component Analysis, PCA) or linear judgment analysis (Linear Discriminant Analysis LDA) waits and obtains.At this, definition ${\left(B^H\right)}^T{I}_i^H=x_i^H,$ B wherein ^HBe that high-resolution human face image characteristic of correspondence is extracted matrix, x _i ^HIt is representative's face image I _i ^HProper vector.The character representation that in like manner, can obtain low-resolution image does $x_i^L={\left(B^L\right)}^T{I}_i^L,$ B wherein ^LBe that low resolution facial image characteristic of correspondence is extracted matrix, x _i ^LRepresent low resolution facial image I _i ^LProper vector.Thus, obtain representing that the set of eigenvectors of corresponding high low resolution training image does $X^H={\left\{x_i^H\right\}}_{i=1}^m$ With $X^L={\left\{x_i^L\right\}}_{i=1}^m,$ Wherein m is a number of samples.

In the super-resolution rebuilding part of recognition feature, the proper vector with high low resolution training image is transformed into the CCA subspace earlier.Concrete, for set of eigenvectors X ^H, X ^LDeduct its average x respectively ^HAnd x ^L, obtain the data set of centralization

With

The target of CCA is to be respectively sample set

With

Seek two groups of base vector V ^H, V ^L, make and utilize the variable after base vector shines upon

With

Between related coefficient ρ reach maximum, promptly have:

$\mathrm{\&rho;}=\frac{E\left[C^H{C}^L\right]}{\sqrt{E\left[{\left(C^H\right)}^2\right]E\left[{\left(C^L\right)}^2\right]}}---\left(1\right)$

Obtain maximal value, wherein E [] represents mathematical expectation.

In order to find the solution base vector V ^HAnd V ^L, definition

Be respectively

With

Auto-covariance matrix,

With

Be respectively With

The cross covariance matrix.Calculate ${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="H2009102075622E00021.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=C_{11}^{-1}{C}_{12}{C}_{22}^{-1}{C}_{21},$ $R_2=C_{22}^{-1}{C}_{21}{C}_{11}^{-1}{C}_{12}.$ R ₁And R ₂Proper vector be the base vector V that asks ^HAnd V ^L

Utilize the base vector of asking, with set of eigenvectors X ^H, X ^LTransform to the CCA subspace, obtain corresponding projection coefficient collection $C^H={\left\{c_i^H\right\}}_{i=1}^m,$ $C^L={\left\{c_i^L\right\}}_{i=1}^m,$ That is:

Because X ^H, X ^LHave relevant immanent structure between these two data sets, it is transformed to the CCA subspace, just behind the correlator space, the linear dependence between two data sets is maximum, makes the consistance of two data set inner topology structures strengthen.Thereby in the CCA subspace, projection coefficient collection C _H, C _LThe hypothesis that the preface of data point keeps between the high lower dimensional space neighborhood of satisfied better local linear embedding algorithm, so can utilize the thought of neighborhood reconstruct to try to achieve the recognition feature of test low resolution facial image in high resolution space.

Low resolution people's face test pattern I to input _l, try to achieve its characteristic of correspondence vector x _l, further with x _lTransform to the CCA subspace, obtain c _l, that is:

c _l＝(V ^L) ^T(x _l-x ^L) (4)

According to the thought of neighborhood reconstruct, at C ^LThe middle c that seeks _lK nearest neighbor { c _Li ^L} _I=1 ^K, and ask the reconstruct weights $W^G={\left\{w_i^G\right\}}_{i=1}^K,$ Make

$\mathrm{\&epsiv;}=\left|\right|c_l-\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G{c}_{\mathrm{li}}^L\left|\right|---\left(5\right)$

Reach minimum, and satisfy $\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G=1.$ These weights are applied to C ^HIn with { c _Li ^L} _I=1 ^KCorresponding { c ^Li _H} _I=1 ^K, can reconstruct I _lThe corresponding recognition feature of full resolution pricture in the CCA subspace is:

$c_h=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G{c}_{\mathrm{li}}^H---\left(6\right)$

At identification division, the present invention carries out recognition of face according to the recognition feature of CCA subspace.Utilize characteristic c _hWith $C^H={\left\{c_i^H\right\}}_{i=1}^m$ Employing is carried out Classification and Identification based on the nearest neighbor classifier of L2 norm, and promptly decision function is:

$g_k\left(c_h\right)=\min \left({\left|\right|c_h-c_{\mathrm{ik}}^H\left|\right|}_2\right),$ i＝1，2，...m (7)

C wherein _Ik ^HExpression C ^HIn belong to i sample of k class.

In order to verify validity of the present invention, utilize CAS-PEAL expression people's face picture library and AR face database to experimentize respectively.The gained experimental result respectively with based on the method for Gunturk in the recognizer of bicubic interpolation, the document and utilize the original high resolution facial image to discern the gained discrimination to compare.Wherein, be meant based on the bicubic interpolation recognizer and utilize bicubic interpolation to obtain the high-resolution human face image, extract the characteristic of this high-resolution human face image then and discern the low resolution facial image; Utilize the original high resolution facial image to discern, promptly directly extract the characteristic of this high-resolution human face image and discern.

Utilize CAS-PEAL expression picture library to carry out in the recognition of face experiment, 190 personages of picked at random from original 377 personages' expression picture library, each personage comprise the image of the different expressions of 5 width of cloth.Fig. 2 has provided the lineup's face image through pretreated a certain personage.

In this experimentation, utilize first three width of cloth image of same personage to train, remain two width of cloth images and discern.Divide three kinds of situation to carry out recognition of face according to high low-resolution image size, be respectively: high resolving power training image size is 64 * 64, and the low-resolution image size is 16 * 16; High resolving power training image size is 64 * 64, and the low-resolution image size is 8 * 8; High resolving power training image size is 32 * 32, and the low-resolution image size is 8 * 8.Utilize PCA to extract characteristic, and the high-definition picture proper vector keep preceding 50 dimensions, preceding 30 dimensions of low-resolution image reservation; In the inventive method, the neighborhood number during neighborhood reconstruct is 20.Parameter in the Gunturk method list of references: preceding 50 dimensional feature vectors are selected in the KL conversion, and the greatest iteration number is 7, λ=0.5, and experimentation used parameter in back is identical therewith.

Can find out that by result among Fig. 3 when image resolution ratio to be identified is low, compare other method, the inventive method discrimination is higher, and overall performance is more stable.The inventive method gained discrimination is with to utilize the original high resolution image to discern the gained discrimination very approaching; This is because the present invention carries out super-resolution rebuilding for recognition feature; Can obtain to test the recognition feature of low-resolution image, effectively raise discrimination in the high-resolution features space.

For the performance of multianalysis algorithm of the present invention, utilize LDA to extract face characteristic, for algorithm of the present invention, compare based on bicubic recognizer and the method for utilizing the original high resolution image to discern.Other parameter setting and computation process are identical with top experiment in the experiment.The gained result is as shown in Figure 4.Comparison diagram 3 can be known with Fig. 4, and the inventive method and the method for utilizing original image identification are utilized LDA to extract aspect ratio and utilized PCA extraction characteristic gained discrimination to want high, and this is consistent with theory.

Utilize the parts of images in the AR image library to carry out the recognition of face experiment, select 135 width of cloth facial images in the picture library, each personage selects three width of cloth images, and wherein preceding two width of cloth are used for training, and last width of cloth is used for test.In this group experimentation, high resolving power training image size is 32 * 32, and the low-resolution image size is 8 * 8.The high low resolution facial image of training of listing lineup's thing is as shown in Figure 5 with test low resolution facial image.

In this experiment, utilize PCA to extract the facial image characteristic equally; High resolving power training image size is 32 * 32; The low-resolution image size is 8 * 8; The neighborhood number is 10, and low-resolution image utilizes PCA to keep preceding 30 dimensional feature vectors, analyzes high-definition picture proper vector dimension and changes the influence to experimental result.The result is as shown in Figure 6.

Can be found out by Fig. 6: along with the increase of proper vector dimension, the gained discrimination is all in rising trend at first, is stabilized in gradually then in the interval, and fluctuation range is no more than 0.05; The method that the inventive method gained discrimination result compares Gunturk can reach the stabilized zone faster, and discrimination is higher; When the proper vector dimension greater than 20; The inventive method gained discrimination with utilize original image gained discrimination very approaching; Sometimes also less times greater than utilizing original image gained discrimination; This is that solution space is limited in the high-resolution features space, can obtain to help identified information because the inventive method is utilized the super-resolution rebuilding of recognition feature.

In sum, the present invention is directed to the lower problem of low resolution facial image discrimination, proposed a kind of super-resolution rebuilding of recognition feature that utilizes and obtained the method for low resolution facial image in the corresponding recognition feature of high resolution space.In the super-resolution rebuilding process of recognition feature, the present invention utilizes CCA to obtain the correlator space of high low-resolution image, in this correlator space, utilizes neighborhood reconstruct to obtain testing the recognition feature of low-resolution image in high resolution space then.Experiment shows that it is less that the inventive method gained discrimination is influenced by the image resolution ratio size, and the gained discrimination is higher.

Claims (3)

1. face identification method based on typical correlation analysis spatial super-resolution is characterized in that: comprise following steps:

1) at first; Utilize classical feature extracting method to extract high low resolution training of human face image recognition characteristic; Utilize two groups of recognition features being extracted as training data; According to based on the requirement of canonical correlation analysis algorithm, must shine upon base vector, according to this mapping base vector recognition feature is transformed into based on canonical correlation analysis correlator space;

Described recognition feature is transformed into based on canonical correlation analysis correlator space and may further comprise the steps:

A) establishing high resolving power and the low resolution training of human face image collection removed after the average is respectively $I^H={\left\{I_i^H\right\}}_{i=1}^m=[I_1^H,I_2^H,...,I_m^H],$ $I^L={\left\{I_i^L\right\}}_{i=1}^m=[I_1^L,I_2^L,...,I_m^L],$ Wherein Be i panel height image in different resolution,

Be i width of cloth low-resolution image, m is the total number of samples of training image; Definition

B wherein ^HBe that the high-definition picture characteristic of correspondence is extracted matrix,

Be to represent facial image

Proper vector; The character representation that in like manner, can obtain low-resolution image does

B wherein ^LBe that low resolution facial image characteristic of correspondence is extracted matrix,

Represent the low resolution facial image

Proper vector; Thus, obtain representing that the set of eigenvectors of corresponding high low resolution training image does With

Wherein

Be i panel height resolution facial image characteristic of correspondence vector,

It is the proper vector of i width of cloth low resolution facial image;

B) for set of eigenvectors X ^H, X ^LDeduct its average respectively

With Obtain the data set of centralization

With

Definition

Be respectively

With

Auto-covariance matrix,

With

Be respectively

With

The cross covariance matrix, wherein E [] represents mathematical expectation, T represents the transposition computing; Calculate

R ₁And R ₂Proper vector be the base vector V that asks ^HAnd V ^L

C) utilize the base vector of asking, with set of eigenvectors X ^H, X ^LTransform to based on the canonical correlation analysis subspace, obtain the corresponding projection coefficient collection of high resolving power and low resolution training facial image

That is:

With

2) secondly, in the correlator space, utilize neighborhood reconstruct to try to achieve the test low resolution facial image recognition feature recognition feature corresponding, realize the super-resolution rebuilding of recognition feature in the high resolving power territory;

3) last, utilize nearest neighbor classifier based on the L2 norm, carry out the arest neighbors Classification and Identification according to what try to achieve based on the recognition feature in the canonical correlation analysis space, thereby obtain discrimination.

2. the face identification method based on typical correlation analysis spatial super-resolution as claimed in claim 1 is characterized in that: said recognition feature super-resolution rebuilding step is following:

1) the low resolution people's face test pattern I to importing _l, try to achieve its characteristic of correspondence vector x _l, further with x _lTransform to based on the canonical correlation analysis subspace, obtain c _l, that is:

2) according to the thought of neighborhood reconstruct, at C ^LThe middle c that seeks _lK nearest neighbor

Wherein

Be c _lI neighbour, and ask the reconstruct weights

Wherein Be the corresponding weights coefficient of i neighbour, make objective function $\mathrm{\&epsiv;}=\left|\right|c_l-\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G{c}_{\mathrm{Li}}^L\left|\right|$ Reach minimum, and satisfy $\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_i^G=1;$

3) with weights

Be applied to C ^HIn with

Corresponding

Promptly reconstruct I _lThe recognition feature of corresponding full resolution pricture is:

3. the face identification method based on typical correlation analysis spatial super-resolution as claimed in claim 1 is characterized in that: said classical feature extracting method adopts principal component analysis (PCA) or linear judgment analysis.

Images