CN103268484A - A Classifier Design Method for High Accuracy Face Recognition - Google Patents

Abstract

The invention provides a design method of a classifier for high-precision face recognition. The design method comprises the following steps of: (1) inputting a vector type face image data set for standardized processing to obtain a face image sample set, wherein the human face image data set must contain the known classes of face image samples and can contain unknown classes of face image samples; (2) utilizing the L1-minimization algorithm to calculate the sparse representation or sparse coding of each face image sample reconstructed by face image samples except the sample; and (3) utilizing the sparse representation or sparse coding of face image samples and the classification information of the classified face image samples to build the optimal model of the classifier, and solving the regularization optimization problem to obtain a classification function. The design method provided by the invention has an explicit way of expression, so that the real-time performance in face recognition application is obviously improved, and the high-precision face recognition is realized under the condition that an image has a great number of noise pixels.

Description

Classifier design method for high-precision recognition of face

Technical field

The invention belongs to field of machine vision, and in particular to a kind of to be used for the classifier design method of high-precision recognition of face.

Background technology

Recognition of face is one of most important research topic in pattern identification research field, is all unusual active research direction both at home and abroad.It is used as a kind of typical biometrics identification technology, it has special advantage in terms of availability, naturality, cost and has been obtained for extensive approval, such as People's Bank of China provides that all national treasury security protections will install face identification system, and the middle Ke Aosen face identification systems of Chinese Academy of Sciences's plum green grass or young crops group research and development are used for Opening Ceremony of the Games etc..

Performance best face identifying system compares cooperation in user in the world at present, and the requirement being normally applied can be met in the case that IMAQ condition is more stable.It is ripe that this is not meant to that face recognition technology has developed into, just the opposite, the need in terms of development and public safety with closed monitor system (CCTV), a greater amount of face recognition application is needed extensive face database, image capture environment be uncontrollable, user it is ill-matched in the case of use, hydraulic performance decline is very fast in this case for existing identifying system, it is impossible to reach realistic scale.The reason for causing this phenomenon mainly has：1. it is difficult to accurate detection and locating human face under uncontrollable environmental condition, 2. it is decreased obviously for the face specimen discerning performance being not covered with training set, 3. complicated illumination causes the apparent violent change of image, 4. the attitudes vibration of identified target, 5. low-quality image problem, the quality of human face image obtained in special occasions (field such as anti-terrorism, criminal investigation) is often excessively poor, in terms of showing as obscuring, there is low noise jamming, resolution ratio, image missing and block, high dimensional data classification problem of 6. magnanimity etc..

Traditional semisupervised classification method can using a small number of samples with classification information and much the samples without classification information come structural classification device (referring to M Belkin, P Niyogi, and V Sindhwani.Manifold Regularization：A Geometric Framework for Learning from Labeled and Unlabeled Examples.Journal of Machine Learning Research, 2006,7：2399-2434).User is when using the grader, only need to specify some positive classification samples and negative classification sample, then obtaining many nothings by computer random sampling specifies the sample of classification information to carry out the construction of grader, and at the same time, the effect of grader can be significantly improved.But existing sorting technique based on the manifold to data distribution it is assumed that validity of the hypothesis on facial recognition data collection does not obtain testing fully confirming.Recently, represent that the algorithm for carrying out face shows excellent Stability and veracity based on Sparse, and the influence to picture noise shows that preferably robustness is (referring to J.Wright, A.Yang, A.Ganesh, S.S.Sastry, Y.Ma, Robust Face Recognition via Sparse Representation, IEEE Transaction on Pattern Analysis and Machine Intelligence 31 (2009)：210-227. and Wright, Y.Ma, J.Mairal, GSpairo, T.Huang, S.Yan, Sparse Representation for Computer Vision and Pattern Recognition, submitted to the Proceedings of the IEEE (2009)).But, represented that algorithm was all offline computational methods based on Sparse in the past, its optimization that complexity will be carried out to each data point is calculated, and it is impossible to meet the requirement of visual information identification in terms of real-time.

The content of the invention

It is used for the classifier design method of high-precision recognition of face the invention provides a kind of, it is therefore intended that the geological information of classified human face data and unfiled human face data can not be utilized simultaneously well by solving existing face recognition technology；Existing rarefaction face recognition algorithms do not have dominant expression simultaneously, it is impossible to the problem of carrying out online recognition of face in real time.

The technical solution adopted by the present invention is：

It is a kind of to be used for the classifier design method of high-precision recognition of face, it is characterised in that to comprise the following steps：

(1) the face image data collection of input vector form, and each facial image sample is standardized, obtain facial image sample set；The face image data, which is concentrated, must include the facial image sample of known class, can include the facial image sample of unknown classification；

(2) L1- norm minimum algorithms are utilized, rarefaction representation or sparse coding that each facial image sample is reconstructed by the facial image sample in addition to the sample is calculated；

(3) optimal model on grader is set up using the rarefaction representation or sparse coding of facial image sample and the classification information for facial image sample of having classified, classification function f (x) is obtained by solving regularization optimization problem；

(4) to the facial image of arbitrary unknown classification information, facial image is converted into vector form with step (1) identical method first, and be standardized, classified using the obtained dominant grader g (x) of classification function f (x), wherein

C is the class number in facial image data set.

Further, step (1) includes following sub-step：

(2.1) for each facial image sample, by its corresponding digital image matrix, in the way of unified row pixel is piled up or row pixel is piled up, a column vector being made up of the pixel value of image is converted into, and the sample vector is subjected to the unitization of mould；

(2.2) data set that the facial image of multiple known category informations is constituted, D × (l+u) rank matrix X=[x for obtaining being made up of image pattern vector are handled by previous step₁..., x_l, x_l+1..., x_l+u], wherein D represents the number of pixels of single image in set, and l represents classification samples number, l ＞ 0；U represents non-classified sample number, u >=0, x_iRepresent the sample vector of some image.

Further, the step (2) is realized by following steps：

Make α_i=(α_1i... α_i-1i0, α_i+1i..., α_{U+l, i})^TCorrespond to sample x for be asked_iBy the reconstruction coefficient vector of other sample rarefaction representations, α is calculated using following L1- norm minimums method_i： ${\mathrm{\α}}_i^{*}=\arg \underset{{\mathrm{\α}}_i,e_i}{\min }\{{\left|\right|{\mathrm{\α}}_i\left|\right|}_1+{\left|\right|e_i\left|\right|}_1\},$ s.t.x_i=X α_i+Ie_i, (1) wherein I is unit matrix, note

Sample x is calculated for formula (1)_iBy the reconstruction coefficient vector of other sample rarefaction representations, then the matrix that the rarefaction representation coefficient vector of training sample is constituted can be designated as $A={[{\mathrm{\α}}_1^{*},{\mathrm{\α}}_2^{*},\·\·\·,{\mathrm{\α}}_{u+l}^{*}]}^T.$

Further, the step (3) includes following sub-step：

(3.1) assume that recognition of face classification function has general expression-form

$f\left(x\right)=\underset{i=1}{\overset{u+l}{\mathrm{\Σ}}}{b}_i{k}_{\mathrm{\σ}}\left({x,x}_i\right)+\mathrm{be},---\left(2\right)$

Wherein b_i=(b_1i..., b_{C, i})^TIt is coefficient vector to be asked, b is departure to be asked, e=(1 ..., 1)^TIt is the column vector of a C dimension, k_σ(x, y) is kernel function, and σ is the parametric variable in kernel function；

(3.2) classification function coefficient matrix B=[b to be asked are remembered₁, b₂..., b_u+l], set up following optimization problem B：

$B^{*}=\underset{B\∈R^{C\×(u+l)}}{\arg \min }\{\frac{1}{l}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{(f\left(x_i\right)-y_i)}^2+{\mathrm{\γ}}_A{\left|\right|f\left|\right|}_K^2+{\mathrm{\γ}}_S{\left|\right|f\left|\right|}_S^2\}---\left(3\right)$

Wherein

It is complexity metrics of the function f (x) in function space,

It is margin of error when function f (x) keeps the rarefaction representation of training sample, γ_AAnd γ_SAll it is previously given arithmetic number, it is B to obtain the solution that discriminant function coefficient table reaches^*, then calculate amount of deflection b^*, finally trying to achieve classification function f (x) is $f\left(x\right)=\underset{i=1}{\overset{u+l}{\mathrm{\Σ}}}{b}_i^{*}{k}_{\mathrm{\σ}}(x,x_i)+b^{*}e.$

The present invention has the advantages that：

(1) present invention is because used the method for sparse coding to describe relation between sample, and the present invention needed the contiguous range parameter of manual intervention algorithm to select unlike former manifold learning；

(2) reconstruction coefficients are calculated present invention utilizes the rarefaction representation of digital image training sample, and grader has been given a definition to reconstructing sparse holding regular terms in Regularization Theory framework so that the discriminant information included in face image data rarefaction representation is more efficiently utilized in our algorithm；

(3) present invention is compared with the classifier methods represented before using Sparse, our invention has dominant expression way, so that its real-time in face recognition application is significantly improved, and high-precision recognition of face can be realized in the case where image has much noise pixel.

Brief description of the drawings

Fig. 1 is used for the flow chart of the classifier design method of high-precision recognition of face for the present invention；

Fig. 2 is some sample images in YaleB face recognition databases；

Fig. 3 is experimental result compares figure；

Fig. 4 a are the facial images of noise-less pollution；

Fig. 4 b are the facial image samples after the noise pollution that variance is 0.05 σ；

Fig. 4 c are the facial image samples after the noise pollution that variance is 0.1 σ.

Embodiment

Inscribed in technical scheme that the invention will now be described in detail with reference to the accompanying drawings between each involved details.It should be noted that described embodiment is intended merely to facilitate the understanding of the present invention, and any restriction effect is not played to it.

As shown in figure 1, being used for the classifier design method of high-precision recognition of face the invention provides a kind of, comprise the following steps：

(1) input face image data and face image data is standardized.It is required that the facial image sample of known identities must be included in this face image data, the facial image sample of substantial amounts of unknown identity can be included；

It is specific as follows：

Input data.

Arbitrary sample x_iOr x_l+jIt is the data of the column vector form changed into by digital picture, y_iRepresent sample x_iClassification information.Wherein l (can not be 0) represents the number of classification samples, and u (can be 0) represents the number of unfiled sample.Sample data x_iWith corresponding categorization vector y_iConcrete form be explained as follows：

For the digital picture of face, a resolution ratio includes m × n pixel for m × n digital picture, represents color comprising 3 numerical informations again on each pixel (the green B of the red G of R are blue).By extracting each row of image array (according to from left to right, the order blue green B of the red G of R) successively, and all row head and the tail tiling methods are constituted vectorial.Size can change into a column vector for having the element of 3 × m × n for m × n color digital image sample.To every by picture pile up after sample x_i, we also have the standardisation process for carrying out that to it mould is 1, i.e. assignment againConversion.So the sample point x used in our graders_iForm be all to have 3 × m × n element, and the column vector of mould a length of 1.Remember the input dimension that D=3 × m × n is data.

Assuming that the class number in face image data is includes some facial images of C known identities in C classes, i.e. data, as sample x_iWhen belonging to pth class, classification information y_i=(0 ... 0,1,0 ... 0)^TIt is to have C element, p-th of element value is 1, remaining is all 0 column vector.

(2) L1- norm minimum algorithms are utilized, the rarefaction representation that each facial image sample is reconstructed by the facial image sample in addition to the sample, or sparse coding is calculated；

It is specific as follows：

For each sample x_i, make D_i=[x₁..., x_i-1, 0, x_i+1..., x_u+l, I], wherein I is the unit matrix of a D rank, then D_iIt is the data matrix of D rows (u+l+D) row；α_i=(α_1i... α_i-1i, 0, α_i+1i..., α_u+l)^TIt is x_iRarefaction linear expression to be asked；Make vectorial θ to be solved_i=(α_1i... α_i-1I0, α_i+1i..., α_u+li, e_1i..., e_Di)^T, then for solving sample data x_iThe L1- norm minimum problems of rarefaction representation be configured to

s.t.x_i=D_iθ_i(specific solution visible document E.Candes and J.Romberg, L1-magic：Recovery of sparse signals via convex programming, http://www.acm.caltech.edu/11magic/, 2005) try to achieve

Its preceding u+l row element is sample data x_iRarefaction representation ${\mathrm{\α}}_i^{*}={({\mathrm{\α}}_{1i}^{*},\·\·\·{\mathrm{\α}}_{i-1i}^{*},0,{\mathrm{\α}}_{i+1i}^{*}\·\·\·,{\mathrm{\α}}_{u+l}^{*})}^T,$ Remember matrix $A={[{\mathrm{\α}}_1^{*},{\mathrm{\α}}_2^{*},\·\·\·,{\mathrm{\α}}_{u+l}^{*}]}^T.$

(3) utilize the rarefaction representation of training data and the classification information of classification based training data sets up the optimal model on grader, classification function f (x) is obtained by solving regularization optimization problem；

It is specific as follows：

Face image data collection with classification information (i.e. identity information) is obtained by step (1)

The rarefaction for obtaining training sample data by step (2) represents sparse vector

Following process is exactly to represent to construct discriminant function f (x) using these sample datas and rarefaction.

Assuming that our discriminant function has general expression-form

Wherein b_i=(b_1i..., b_{C, i})^TIt is coefficient to be asked, b is departure to be asked, e=(1 ..., 1)^TIt is the column vector of a C dimension, k_σ(x₁, x₂)=exp (- | | x₁-x₂||²/σ²) it is given gaussian kernel function, σ is empirical parameter variable manually given in the kernel function, and the selection of its occurrence can also use well-known cross validation method.Remember B=[b₁..., b_u+l] it is grader sparse matrix to be asked.

In order to utilize the face sample with classification information, while the grader of a high recognition performance is constructed using the rarefaction representation information without classification information face sample, it is proposed that solving following problem carrys out structural classification device

$B^{*}=\underset{B\∈R^{C\×(u+l)}}{\arg \min }\{\frac{1}{l}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{(f\left(x_i\right)-y_i)}^2+{\mathrm{\γ}}_A{\left|\right|f\left|\right|}_K^2+{\mathrm{\γ}}_S{\left|\right|f\left|\right|}_S^2\}---\left(1\right)$

Wherein

It is complexity metrics of the function f (x) in functional Hilbert spaces,It is the error metrics that function f (x) keeps producing during the rarefaction representation of training sample, γ_AAnd γ_SAll it is the artificial arithmetic number empirically given, the selection of its occurrence can also use well-known cross validation method.

Known K is that the element on (l+u) × (l+u) matrix, the i-th row j column positions is k_σ(x_i, x_j)；

It is rarefaction representation matrix；Y=[y₁..., y₁, 0 ..., 0] and it is the classification information matrix that C rows (u+l) are arranged；J=diag (1 ..., 1,0 ..., 0) is the diagonal matrix that size is (u+l) × (u+l), and its preceding l diagonal element value is 1, and remaining is 0.We define rarefaction coding keep error metrics be

${\left|\right|f\left|\right|}_S^2=\frac{1}{{(u+l)}^2}\underset{i=1}{\overset{u+l}{\mathrm{\Σ}}}{\left|\right|f\left(x_i\right)-\underset{j=1}{\overset{u+l}{\mathrm{\Σ}}}{\mathrm{\α}}_{\mathrm{ji}}^{*}f\left(x_j\right)\left|\right|}^2$

Theoretical according to reproducing kernel Hilbert space, the complexity metric of function in space is

${\left|\right|f\left|\right|}_K^2=\mathrm{tr}\left({\mathrm{BKB}}^T\right)---\left(3\right)$

Then the optimization problem that our discriminant function is solved, formula (1) is converted into following matrix form problem

$B^{*}=\underset{B\∈R^{C\×(u+l)}}{\arg \min }\{\frac{1}{l}\mathrm{tr}\left((Y-\mathrm{JKB}){(Y-\mathrm{JKB})}^T\right)$

$+{\mathrm{\γ}}_K\mathrm{tr}\left({\mathrm{BKB}}^T\right)+\frac{{\mathrm{\γ}}_S}{{(u+l)}^2}\mathrm{tr}\left(\mathrm{BK}{(I-A)}^T(I-A){\mathrm{KB}}^T\right)\}---\left(4\right)$

Problem (4) can be by object function, seeking B local derviation and making function after its derivation be equal to 0, you can obtains the least square solution of optimization problem (4)

$B^{*}=Y{(\mathrm{KJ}+{\mathrm{\γ}}_K\mathrm{lI}+\frac{{\mathrm{\γ}}_{S}l}{{(u+l)}^2}K{(I-A)}^T(I-A))}^{-1}---\left(5\right)$

Then the amount of deflection is calculated

The wherein F of matrix₂Then the quadratic sum of norm elder generation calculating matrix all elements carries out evolution to it.The discriminant function finally tried to achieve is expressed as

$f\left(x\right)=\underset{i=1}{\overset{u+l}{\mathrm{\Σ}}}{b}_i^{*}{k}_{\mathrm{\σ}}(x,x_i)+b^{*}e.$

(4) to the facial image sample of arbitrary unknown identity information, first with and the first step identical method image is turned into vector form and be standardized, classified using the obtained dominant grader g (x) of function f (x).

It is specific as follows：

Obtain after discriminant function f (x), for the facial image sample x of arbitrary unknown identity information, the method for being first according to the first step is translated into vector form, and sample vector then is carried out into mould standardization

Sample x is substituted into discriminant function f (x), the output of discriminant function is a multivalue vector, might as well set f (x)=(f₁(x) ..., f_C(x))^T, then x classification g (x) judge can obtain with the following method

$g\left(x\right)=\arg \underset{i=1,\·\·\·,C}{\max }\left\{f_i\left(x\right)\right\}$

Can be widely used in using the dynamic vision system of the inventive method needs the place of recognition of face, intelligent transportation, military, customs, bank, hotel, enterprise, and department gateway such as government etc. needs to carry out the place of Automatic face recognition.Especially in the case that the quality of human face image of acquisition is low, the effect of identification is especially pronounced compared to analogous algorithms.

We used YaleB face recognition databases.Such as Fig. 2, the database contains the 2114 width human face photos of 38 people, and each width human face photo is all under different illumination conditions 32 × 32 gray level image.Everyone probably correspond to 60 human face photos in the database.We are random from database to select 15% data as sightless test data during training, in remaining 85% training data, we give m sample class information (i.e. identity information) in every class data, and remaining is used as the training data without classification information (identity information).Specific classifier parameters of the present invention are set to, but are not limited to, kernel function k_σ(x₁, x₂)=exp (- | | x₁-x₂||²/σ²) in σ=0.5, regularization parameter λ_A=0.005 and λ_S=0.01.

What Fig. 3 was provided is the experimental result of the invention with face recognition algorithms leading in the world on the database, and wherein abscissa represents the number for having classification information facial image in every class, and ordinate represents accuracy rate of the algorithm in training stage sightless test data.In figure, our algorithmic notation is S-RSLC, general arest neighbors sorting technique in 1-NN algorithmic notation recognitions of face, GFHF is document X.Zhu, Z.Ghahramani, J.Lafferty, Semi-supervised learning using Gaussian fields and harmonic functions, in：The 20th International Conference on Machine Leaming (ICML), 2003, pp.912-919. the method in, SRC is document J.Wright, Y.Ma, J.Mairal, G.Spairo, T.Huang, S.Yan, Sparse Representation for Computer Vision and Pattern Recognition, method in submitted to the Proceedings of the IEEE (2009), LapRLSC is document M Belkin, P Niyogi, and V Sindhwani.Manifold Regularization：A Geometric Framework for Learning from Labeled and Unlabeled Examples.Journal of Machine Learning Research, 2006,7：Method in 2399-2434..

For stability of the verification algorithm to image pixel noise, all pixels in each image in database are all added independent identically distributed Gaussian noise by us.We generate two noise databases with two kinds of noises.One storehouse is that, by adding a variance for 0.05 σ, average produces for 0 Gaussian noise to pixel；One storehouse is that, by adding a variance for 0.1 σ, average produces for 0 Gaussian noise to pixel, and wherein σ is the variance yields of distance between all samples.Fig. 4 a are the facial images of noise-less pollution, and Fig. 4 b are the facial image samples after the noise pollution that variance is 0.05 σ, and Fig. 4 c are the facial image samples after the noise pollution that variance is 0.1 σ.As can be seen that people can hardly recognize the facial image of Noise.But the conclusion that Tables 1 and 2 is provided is shown, our method can very stably recognize the identity of Noise facial image.This is significantly to recognition of face problem.

Table 1 is recognition of face effect of all comparison algorithms on noise contaminated data collection (without classification results on classification information sample in training set)

Recognition of face effect (training stage sightless test data in face recognition result) of all comparison algorithms of table 2 on noise contaminated data collection

It is described above; embodiment only in the present invention; but protection scope of the present invention is not limited thereto; it is any be familiar with the people of the technology disclosed herein technical scope in; it is appreciated that the conversion or replacement expected; it should all cover within the scope of the present invention, therefore, protection scope of the present invention should be defined by the protection domain of claims.

Claims (4)

1. a classifier design method for high precision face recognition, characterized in that, comprising the steps: (1) Input a face image dataset in the form of a vector, and standardize each face image sample to obtain a face image sample set; the face image dataset must contain face image samples of known categories, which can be Contains face image samples of unknown categories; (2) use

-a norm minimization algorithm, calculating the sparse representation or sparse coding of each face image sample reconstructed by face image samples other than the sample; (3) Use the sparse representation or sparse coding of the face image samples and the category information of the classified face image samples to establish an optimal model for the classifier, and obtain the classification function by solving the regularization optimization problem ; (4) For any face image with unknown category information, first use the same method as step (1) to convert the face image into a vector form, and perform standardized processing, using the classification function

The resulting explicit classifier

to classify, among them,

,

is the number of categories in the face image dataset. 2. The classifier design method for high-precision face recognition according to claim 1, wherein step (1) includes the following sub-steps: (2.1) For each face image sample, convert its corresponding digital image matrix into a column vector composed of the pixel values of the image according to the uniform row pixel stacking or column pixel stacking method, and perform this sample vector Unitization of modules; (2.2) A data set composed of multiple face images with known category information, obtained through the previous step processing and composed of image sample vectors

order matrix

,in Indicates the number of pixels of a single image in the collection,

Indicates the number of classified samples,

; Indicates the number of unclassified samples,

, A vector of samples representing an image. 3. The classifier design method for high-precision face recognition according to claim 2, wherein said step (2) is realized by the following steps: make

corresponding to the samples to be sought The reconstructed coefficient vector sparsely represented by other samples adopts the following - Norm minimization method calculation

: st

, (1) in

is the identity matrix, remember

The sample calculated for formula (1)

The reconstructed coefficient vectors are sparsely represented by other samples, then the matrix formed by the sparsely represented coefficient vectors of the training samples can be written as . 4. The classifier design method for high-precision face recognition according to claim 3, wherein the step (3) includes the following sub-steps: (3.1) Assume that the face recognition classification function has a general expression form

, (2) in

is the coefficient vector to be sought,

is the desired deviation,

Is a

column vector of dimensions,

is the kernel function, is the parameter variable in the kernel function; (3.2) Remember the coefficient matrix of the classification function to be found

, establish the following optimization problem to solve :

(3) in

is a function

A complexity measure in function space, is a function

The amount of error when maintaining a sparse representation of the training samples, and

are all pre-given positive real numbers, and the solution expressed by the coefficients of the discriminant function is

, and then calculate the deviation value , and finally get the classification function for .

Images