The present invention relates to the field of the recognition of faces in digital images.
In the field of face recognition, use is made nowadays of statistical procedures for classification by reduction of dimensions. Statistical models of reduction of dimension are constructed by virtue of a set of images of faces that is referred to hereinafter as the “learning base” (or BA for short). One and the same person is preferably represented several times in the learning base. The various representations of the face of one and the same person are then referred to as a “class”.
The two major applications of face recognition are “authentication” and “Identification”.
Within the framework of authentication, the face to be recognized is compared with the faces of the learning base so as to assign it an identity. Biometry or video surveillance, for example, are applications based on authentication.
Within the framework of identification, once the model has been constructed, the learning base is no longer used. One has a first unknown face and a second unknown face and one seeks to compare these two faces so as to determine whether or not it is the same person.
The indexing of movie sequences, for example video sequences, may use applications of this type. The interest could for example be to detect in a video all the sequences where one and the same person is found.
The recognition of faces in digital images currently proposes various methods that can be classed in the following manner:
Tools for statistical analysis of data of large dimension n make it possible to reduce a complex system of correlations by projecting initial data into a space of lesser dimension k<<n. According to the intended objective, the projection space thus obtained may give:
Described hereinafter is the procedure termed “principal component analysis” of the prior art and set forth in particular in:
“Eigenfaces for recognition”, M. Turk and A. Pentland, Journal of Cognitive Neuroscience, vol. 3, March 1991, pages 71-86.
A commonly used definition of the PCA is that which associates, with an input set of vectors, a set of orthogonal principal axes (termed “principal components”) and on which the projection of the variance of the input vectors is a maximum.
Here, it is indicated that the term “vector” refers to a column vector. Moreover, X denotes the vector to be projected, of large size l·m. Its orthogonal projection is denoted {circumflex over (X)}. This projection is performed onto the space of dimension k, an orthogonal base of which is stored in the form of columns in the matrix P. The matrix P is therefore of size (l·m)×k. The projection of the vector X is then expressed by:
{circumflex over (X)}=PTX (1)
The matrix P is called the “matrix of projection of the initial space into the space of principal components”.
As criterion to be maximized we choose:
J(P)=trace(Sp) (2)
where Sp denotes the covariance matrix of the learning base projected into the base of P, i.e.:
where n denotes the number of images present in the learning base.
If Xi denotes the input vector corresponding to the ith vector of the learning base BA, we have:
It is indicated that maximizing the criterion according to relation (2) amounts to maximizing the criterion:
where is the set of matrices with real coefficients of size (l·m)×k.
S is the covariance matrix of the learning base BA, of size (l·m)×(l·m), given by the following relation:
where, with the notation given previously:
It is shown that, under the hypothesis that the vectors Xi of the learning base are Gaussian vectors, that are pairwise independent and identically distributed (property denoted “iid” hereinafter), P is composed of the k eigenvectors of S that are associated with the k largest eigenvalues (k fixed).
The PCA is commonly used to represent or recognize faces. The procedure for recognizing faces that is best known and based on the PCA has been proposed in the aforesaid document: “Eigenfaces for recognition”, M. Turk and A. Pentland, Journal of Cognitive Neuroscience, vol. 3, March 1991, pages 71-86.
The procedure requires a learning base consisting of a set of images presented as input in the form of a vector per image. Each image Xi, consisting of l rows and m columns of pixels as grey levels, is thus reduced to a vector of size l·m by concatenating its rows of pixels. A PCA is performed directly on these vectors, giving a set of k principal components of the same size l·m as the initial image vectors and are designated by the term “eigenfaces”. The number k of principal components to be retained may be fixed or else determined from the eigenvalues.
The comparison between two images of faces is made following a projection into the base of the eigencomponents according to relation (1) hereinabove. The two projected vectors are compared according to a measure based on a predetermined criterion of similarity.
It is shown in particular that the principal components constitute the subspace of dimension k minimizing the mean quadratic error of reconstruction, defined as being the distance denoted L2 between the learning base and its orthogonal projection in the base consisting of the principal components.
However, a drawback of the procedure resides in the fact that this base does not necessarily offer an optimal classification of the data. Specifically, the principal components maximize the total variance of the learning base, without distinguishing the variations internal to each class from the variations between classes.
Described hereinafter is a procedure arising out of Linear Discriminant Analysis (or LDA), of the prior art, commonly used in shape recognition and described in particular in:
“Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection”, P. Belhumeur, J. Hespanha and D. Kriegman, Special Theme Issue on Face and Gesture Recognition of the IEEE Trans. on Pattern Analysis and Machine Intelligence, 19(7), pages 711-720, July 1997.
It makes it possible to find the subspace that jointly maximizes the variations between classes and minimizes the mean variation inside the classes, so as to give a subspace called the “discriminant subspace” of the initial space and ensuring better discrimination between classes.
The LDA is distinguished from the PCA in particular in that the LDA is a supervised procedure, that is to say the construction of the model requires, for each image of the learning base, the datum of a vector as well as the datum of its membership class.
Here the hypothesis is made that the images of one and the same class, these being represented in the form of vectors, are Gaussian vectors having the iid property described hereinabove. P denotes the orthogonal projection matrix from the initial space into the discriminating subspace within the sense of equation (1) hereinabove. If we choose as criterion to be maximized the expression:
J(P)=trace((Swp)−1Sbp) (6)
where Swp (respectively Sbp) is the intra-class (respectively inter-class) covariance matrix of the learning base projected by the orthogonal projection P. P is a matrix of size (l·m)×k. It is recalled that a class is a set of representations of one and the same face in the learning base. It is indicated also that the terms “intra-class” are concerned with properties within one and the same class, whereas the terms “inter-classes” are concerned with properties from one class to another.
If C denotes the total number of classes of the learning base, we define:
where:
It is then indicated that maximizing the criterion according to relation (5) amounts to choosing the matrix P in the following manner:
where:
The columns of P contain the k eigenvectors of the matrix Sw−1Sb associated with the k largest eigenvectors, where Sw−1 is the inverse of Sw.
As a general rule, the dimension of the input vectors is much bigger than the number of examples acquired in the learning base (l·m>>n). The matrix Sw is then singular and noninvertible.
It is then possible to perform the LDA in a base, determined previously by applying the PCA procedure, the dimension of this base being less than the number of examples of the learning base. This approach is designated hereinafter by the abbreviation “PCA+LDA”.
The number of components to be retained, called “Fisherfaces”, may be determined thereafter in the same manner as for the PCA described hereinabove. The classification is performed after orthogonal projection into the space of Fisherfaces, in the same manner as for the PCA.
Following the numerous comparative studies between the PCA and LDA of the prior art, it is taken as read that if one considers a learning base of sufficient size and sufficiently representative, the LDA gives better results than the PCA and makes it possible above all to better manage the differences of illumination in the pictures of the faces, the differences of facial expressions and of pose.
Very recently, another procedure based on a “bidimensional” PCA has been proposed in:
“Two-dimensions PCA: A New Approach to Appearance-Based Face Representation and Recognition”, J. Yang, D. Zhang, A. F. Frangi and J. Y. Yang, in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26, no. 1, January 2004.
This procedure is described hereinafter.
One constructs a model on the basis of the set of images of the learning base that are stored in the form of matrices of pixels having l rows and m columns. More particularly, we seek k orthonormal vectors, P=[Pl, . . . , Pk] of length m, such that the projection of the learning base onto this base of vectors ensures a maximum of variance of the learning base. The projection of an image X of size l×m onto the matrix P of size m×k is given by the following linear relation, where {circumflex over (X)} is a matrix of size l×k:
{circumflex over (X)}=XP (11)
The criterion to be maximized is as follows:
J(P)=trace(Sp) (12)
Sp designates the covariance matrix (termed “bidimensional” as opposed to a unicolumn vector) of the n vectors of the images projected from the learning base onto the vector base P. If we consider:
In this expression, S is the covariance matrix of the columns of the images and it is calculated as follows:
where
is a mean matrix, of size l×m, of the n images of the learning base.
The criterion to be maximized according to relation (14) is called the “generalized total dispersion criterion”. As for the PCA, the k vectors └Pl, . . . , Pk┘ to be retained are the eigenvectors of the matrix S corresponding to the largest eigenvalues. P=└Pl, . . . , Pk┘ denotes the projection matrix within the sense of relation (11). The projection of the image Xl by P is denoted {circumflex over (X)}i=└{circumflex over (X)}il, . . . , {circumflex over (X)}ik┘ where {circumflex over (X)}il=XiPj is the vector of length l (projection of the image Xi onto the vector Pj).
The number k of components to be retained may be determined in the same manner as for the PCA, seen hereinabove.
As before, the comparison between two faces is performed in the projected space. Here, the projection of an image onto P no longer gives a vector, but a matrix. Therefore, here we use a measure of similarity between matrices {circumflex over (X)}i of size l×k. The distance between the matrices {circumflex over (X)}i and {circumflex over (X)}t may be as follows:
The bidimensional PCA, conducted with a very small number of components k, gives better results than the classical PCA, on the same bases “Yale Face DataBase B” “ORL” and “AR”, which are among the bases of faces that are best known.
As indicated hereinabove, numerous procedures for reducing dimensions having however been proposed in the prior art, the major disadvantage of the PCA and LDA procedures resides in the fact that the size m·l of the vectors of the learning base is generally very large, this leading to:
Within the framework of the LDA, to circumvent this problem, one generally performs two projections instead of just one, by combining the PCA and LDA procedures (so-called “PCA+LDA” processing). However, this approach considerably increases the complexity of the LDA. Moreover, the choice of linear combinations of the principal components as discriminating components, on the one hand, is not justified and, on the other hand, it lacks rigor.
The bidimensional PCA, even though it does not exhibit this drawback and though it guarantees optimal conservation of the global variance of the projection learning base, does not make it possible to distinguish variations between classes and variations inside classes. Like the PCA procedure, this technique is well suited to the reconstruction of images of faces after compression by reduction of dimension. However, as for the PCA, it does not necessarily ensure good discrimination between classes.
The present invention is intended to improve the situation.
An aim of the invention is to guarantee better discrimination between the classes of the learning base.
Another aim of the invention is to afford a reduction in the calculations and in the memory space necessary for storing the model constructed so as to be able to classify the images, with respect to the standard discriminating procedures such as the LDA.
For this purpose it proposes a method of recognizing faces in digital images, by statistical processing, in which there is provided a learning base containing images of faces, this learning base being hierarchized as a plurality of classes each comprising different pictures of one and the same person.
In the method within the meaning of the invention, the images are represented by matrices whose coefficients are values of pixels. The method then comprises a preprocessing of the learning base, in which:
According to one of the advantages afforded by the present invention, the method is more robust to changes of facial expressions than the technique based on the PCA2D of the prior art and its implementation requires much less memory size and less capacity for calculations than the classical LDA technique.
In a preferred embodiment, the preprocessing of the learning base comprises the following steps:
To reduce the dimension of the vector base, the eigenvectors associated with a chosen limited number of largest eigenvalues are advantageously selected.
In another embodiment, the method which is the subject of the invention comprises a prior step to the formation of said matrix system and consisting in applying a transposition of the matrices representing the images of the learning base. It is indicated that this variant leads to discriminating components denoted “2D-rows”, whereas the succession of steps above, without the transposition of the matrices, leads to discriminating components denoted hereinafter “2D-columns”. This distinction will be made explicit in greater detail later.
Advantageously, the method can comprise a subsequent classification processing of an image of a face to be recognized, comprising the following steps:
Typically, said projection is performed by multiplying the matrix associated with the face to be recognized and/or the matrix associated with a reference face, by each of the selected eigenvectors.
In a first embodiment, the distance criterion comprises the evaluation of a sum of the distances between the corresponding columns of the identification and reference matrices. This distance may be given by relation (16) hereinabove but applied within the framework of the invention.
As a variant, the distance criterion is fixed between the identification matrix and the projected mean of each class of the learning base, in the guise of reference matrix. This distance is preferably estimated by relation (24), given and commented on later. Here, it is therefore possible to define a reference image as being the mean image of a class.
The reference image may on the contrary be a real image derived directly from the learning base. In this case, a possible application is the triggering of an alarm of a surveillance system if the comparison between a face to be recognized and one or more reference faces belonging to the learning base exhibits a distance greater than a threshold.
In another variant, the reference image does not belong to the learning base but it is derived from a previous picture. In this case, a possible application may consist in knowing how many times a person appears in a video sequence, a previous image of this person then being the reference image.
The method within the sense of the invention may advantageously comprise an additional step of reconstructing the images of faces, after a compression of the image data by reduction of dimension, since the choice of a limited number of discriminating components (or of eigenvectors as indicated hereinabove) allows this reduction in dimension.
The present invention may be implemented by a computer workstation driven by running an appropriate piece of software (or “computer program product”). In this regard, the present invention is aimed at such a computer program product, intended to be stored in a memory of a processing unit, or on a removable memory medium and intended to cooperate with a reader of said processing unit. This program comprises instructions for the implementation of all or part of the steps of the method hereinabove.
Other characteristics and advantages of the invention will become apparent on examining the detailed description hereinafter, and the appended drawings in which:
and
A preferred embodiment of the present invention is aimed at a method based on a supervised statistical procedure, hereinafter referred to as “Bidimensional Linear Discriminant Analysis” (or “LDA2D”).
The LDA2D can take two forms: LDA2D-columns and LDA2D-rows. These two procedures may optionally be combined so as to cross-check the information and/or compare the results that they each give.
Referring to
These vectors constitute a vector base which will be used during the classification described hereinafter.
It is indicated that a variant of the matrix multiplication step consists rather in calculating the product of the inverse of the 2D-columns intra-class covariance matrix times the 2D-columns inter-class matrix. Next we calculate the eigensystem of this matrix product as previously.
The LDA2D-rows consists of an equivalent method, but performed using the transposed matrices of the 2D images of faces. Here again, we evaluate a matrix product involving the intra-class and inter-class covariance matrices to form a system of discriminating components denoted “2D-rows”.
The general method of processing may, advantageously but not necessarily, be concluded with a step 20 of reconstructing the faces after reduction of dimension (box C of
Of course, the steps of the formal calculation of the LDA2D are given only roughly in
These steps are now described in greater detail. It is indicated that the notation employed hereinafter is then different from that which appears in
The Bidimensional Linear Discriminant Analysis is used as a statistical tool for classifying data. As for the PCA2D described hereinabove, the images of the learning base are represented in the form of real matrices of size l×m. We seek the k vectors P of size m maximizing the criterion according to the relation:
J(P)=tr(Swp
where:
It will be noted that {circumflex over (X)}i, {circumflex over (
It is indicated that maximizing the criterion according to relation (17) amounts to finding the set of components P that jointly maximizes the inter-class variance and minimizes the intra-class variance. Intuitively, it will be understood that the discrimination between projected classes may be significant. We show that:
Swp=PTSwP and Sbp=PTSbP (20)
where:
It is recalled that Xi,
We refer to the matrix Sw (respectively the matrix Sb), of size m×m, as the “2D-columns intra-class covariance matrix” (respectively the “2D-columns inter-class covariance matrix”).
Maximizing the criterion according to relation (17) therefore amounts to finding the vectors P maximizing the ratio:
One is then dealing with the eigenvectors associated with the matrix Sw−1Sb corresponding to the largest eigenvalues.
If we assume that the number l of rows of the images of the learning base is less than the number n of examples present in the learning base (n>>l), this corresponding to a condition which generally holds, the matrix Sw is invertible. Consequently, recourse to a first reduction of dimension, such as implemented in the PCA+LDA procedure described hereinabove, is unnecessary here.
This first form of implementation of the invention is called “Bidimensional Columns Linear Discriminant Analysis” (or “LDA2D-columns”).
Once the model has been constructed (following steps 11 to 17 hereinabove), we therefore have a base P=└Pl, . . . ,Pk┘ of k eigenvectors of dimensions m, allowing good discrimination of the classes of the learning base by projection into this base. These vectors will be called “2D-columns discriminating vectors”.
It is indicated that as a variant, the Bidimensional Linear Discriminant Analysis, within the sense of the invention, may also be done using transposed image-matrices of the learning base. For any image Xi of size l×m, we consider a new mode of projection onto the base of vectors P=└Pl, . . . ,Pk┘, which is now of size l×k and the definition of the projection according to relation (11) is replaced by:
{circumflex over (X)}i=XiTP (22)
The maximization criterion is still based on relation (17) but the definition of {circumflex over (X)}i changes here. {circumflex over (
Sw and Sb are henceforth calculated as follows:
This new matrix Sw (respectively Sb), of size l×l, is called the “2D-rows intra-class covariance matrix” (respectively “2D-rows inter-class covariance matrix”).
The method continues as described above. It may be considered that the number m of columns of the images of the learning base is assumed to be less than the number n of examples present in the learning base (m>>n). The matrix Sw is therefore invertible. The construction of this new model therefore leads to the determination of a base P=└Pl, . . . ,Pk┘ of k vectors of dimensions l, allowing good discrimination of the classes of the learning base by projection into this base. These vectors are called “2D-rows discriminating vectors”.
The latter mode of implementation of the Bidimensional Linear Discriminant Analysis is referred to as “Bidimensional Row Linear Discriminant Analysis” (or “LDA2D-rows”).
The general method (using an LDA2D-rows or an LDA2D-columns) continues with a classification of the faces of the learning base (box B of
Here we present the classification for the LDA2D-columns. However, the transposition to LDA2D-rows is deduced therefrom naturally, by replacing the projection according to relation (11) by the projection according to relation (22).
The matrix of each image Xi of the learning base is projected into the base P=└Pl, . . . ,Pk┘ within the sense of relation (11) to obtain a matrix {circumflex over (X)}i per image, of size l×k, corresponding intuitively to the position of the image Xi in the base P of the 2D-columns discriminating components. To compare two images of faces Xi and Xt in this base, the two matrices {circumflex over (X)}i and {circumflex over (X)}t must be compared. For this purpose, two similarity measures are provided:
Specifically, the LDA2D, unlike the PCA, is a supervised procedure. Hence the membership classes of the images of the learning base are known. Within the framework of an authentication using the LDA2D, it is therefore possible to define a more appropriate distance than the column distance L2, as is given by the above expression (24).
The image of the face to be recognized by Xt will subsequently be considered to belong to the nearest class in the sense of the distance according to formula (24), or else to the same membership class to which the nearest image belongs, according to the column distance L2.
It is indicated that the choice of the number of 2D discriminating components to be retained may be made exactly in the same way as in the framework of a classical PCA or LDA, such as described in particular in: “Analyzing PCA-based Face Recognition Algorithms: Eigenvector Selection and Distance Measures”, W. S. Yambor, B. Draper, and R. Beveridge, 2nd Workshop on Empirical Evaluation in Computer Vision, 2000.
The reconstruction of faces (box C of
According to a property of the LDA2D, the projection {circumflex over (X)}i of the image Xi and the corresponding 2D discriminating components └Pl, . . . ,Pk┘ (rows or columns) may be combined to reconstruct an image of a face in the initial space The eigenvectors └Pl, . . . ,Pk┘ being orthonormal, the reconstruction X′i of the image Xi is given simply by the following relation, using the LDA2D-columns:
X′i=XiPPT={circumflex over (X)}iPT (25)
The reconstruction using the LDA2D-rows will be done according to the equation:
X′i=PPTXi=P{circumflex over (X)}iT (26)
with the appropriate notation of LDA2D-rows.
Described hereinafter is a particular embodiment, with an application to authentication. For this purpose, we construct a learning base as well as a test base, containing faces to be recognized.
The base of Intelligent Multimedia Laboratory (IML), downloadable at the site http://nova.postech.ac.kr/ has been chosen.
Represented in
The base exhibits variations of scale. Occultations such as the wearing of spectacles, the wearing of headgear, or the like, are also represented therein.
The tests conducted used two subsets of this IML base. The first subset makes it possible to illustrate the robustness of the procedures evaluated to pose (subsequently referred to as the “base of poses”). The second subset makes it possible to illustrate the robustness of the procedures evaluated to facial expression (referred to subsequently as the “base of expressions”).
To test the robustness to pose, use is made of a learning base composed of five hundred and thirty-five images of faces (five views per class), according to different poses which are however close to the frontal pose. The test base, for its part, is composed of four hundred and twenty-eight images (four views per person), under more accentuated poses. An extract of the learning base and of the test base is given in
To test the robustness of the method to facial expressions, use has been made of a learning base composed of three hundred and twenty-one images of faces (three frontal views per person). One view represents a neutral facial expression, the other two are chosen at random from among the four facial expressions proposed in the base. The test base, for its part, is composed of two hundred and fourteen images (two views per person), representing the remaining facial expressions.
An extract of the learning base and of the test base is given in
All the images of the learning base, like those of the test base, have been preprocessed in such a way that the images are preferentially of size 150×130 pixels, with a distance of 70 pixels between two centers of pupils. The right eye and the top left corner of the image are situated at a distance of 30 pixels along the abscissa axis and of 45 pixels along the ordinate axis. The images, initially color, are simply transformed into luminance images, having grey levels, without any other preprocessing.
The variables considered are digital images, hence taking the form of discrete bidimensional variables. The processed images are for example sampled over a discrete grid of 150 rows and 130 columns and of values in the real luminance space of between 0 and 255.
The construction of the LDA2D-columns model is now described. Using the images of the learning base that are stored in the form of matrices of 150×130 pixels, we calculate the covariance matrices Sw and Sb given by relation (23). In particular we calculate the eigensystem of Sw−1Sb with the aid of a singular value decomposition. We ascertain the matrix P=└Pl, . . . ,Pk┘ of the 2D-columns discriminating components. From this we deduce by projection by applying relation (11) the projections {circumflex over (X)}i of the faces of the learning base according to the matrix P, that we keep in memory for the following recognition phase.
For the recognition of the faces of the test base, all the faces Xt of the test base are projected into the space P of 2D-column discriminating components by applying the equation (11), to obtain the matrices {circumflex over (X)}t.
It is recalled that all the people an image of whom is in the test base are also represented across different images in the learning base.
The test image will be simply assigned to the class c that is nearest according to the distance criterion chosen. The similarity measures are given by the column distance L2 of relation (16) and by the other distance estimate given by relation (24).
Allocation to a class and reconstruction of the faces are done as described hereinabove.
Once the faces have been reconstructed, they can be displayed in the form of images of size 150×130 pixels.
In what follows, we present a comparison of the experimental results between:
To test the robustness to pose, the results of the LDA2D within the sense of the invention are compared with the results of the classical LDA, calculated according to the combination PCA+LDA procedure. To determine the number of principal components from which the LDA must be constructed, the PCA procedure has first of all been conducted in the base of poses. With more than 200 principal components, the rates of recognition decline substantially. It has therefore been chosen to retain 200 principal components to construct the classical LDA.
The distances used for the classification are a simple vector-to-vector distance L2 (designated by “L2” hereinafter), as well as a distance L2 from a vector to the arithmetic mean of the projected vectors of each class (denoted by “L2moyCl”).
The best recognition rate of the LDA2D for the distance L2 is 94.4% and is achieved with only 8 components. For the distance L2 moyCl, the best recognition rate is 95.8% and is achieved with only 13 components.
As represented in
Thus, the LDA2D-columns give substantially better results than the classical LDA, and especially with a memory space required for the storage of the model which is much less.
Hereinafter, when alluding to the LDA2D, it will be understood that one is dealing with the LDA2D-columns.
Referring now to
Represented in
The results of the LDA2D are now compared with those of the classical LDA and with those of the PCA2D to test the robustness to changes of facial expressions. The processing is applied to the “base of expressions” defined hereinabove.
It is noticeable especially that the LDA2D rows gives a very different classification from the LDA2D columns. In particular, the errors of these two procedures are not systematically made on the same images. Even in the case of errors made by the two procedures on one and the same image, the identity assigned is not systematically the same. The bases of 2D-rows and 2D-columns discriminating components therefore appear to be very different. However, the two procedures offer good results on the two bases tested. As indicated hereinabove, an optional merging and/or cross-checking of the results obtained by the two procedures LDA2D-rows and LDA2D-columns may be provided for.
Comparison of
The results of the LDA2D and of the PCA2D of the prior art are now compared.
We now refer to
These two people are different but bear the same facial expression however. The LDA2D-rows with 5 components, for its part, assigns the image of the face given in
Other tests have shown that in the framework of the PCA2D with 18 components, this type of error occurs 63 times out of 93 errors in all (214 faces in the base of tests). Errors of this type therefore represent 67.74% of the classification errors made by the PCA2D. There remain just 30 errors which are not due to a separation as a function of facial expression.
Within the framework of the LDA2D rows and with 5 components, this type of error occurs 31 times out of 59 errors in all (214 faces in the test base). Errors of this type therefore represent only 52.54% of the classification errors made by the LDA2D, i.e. 25.20% fewer than the PCA2D. There remain 28 errors which are not due to a separation as a function of facial expression. The number of residual errors (which are not due to this confusion between facial expression and identity) is therefore substantially the same as for the PCA2D. It can therefore be considered that the best recognition rate of the LDA2D (11.2% of improvement with respect to the PCA2D) stems directly from the reduction in errors of this type that are due to facial expressions.
The PCA2D therefore appears to be overly influenced by the intra-class variations, with respect to the LDA2D within the sense of the invention. In the PCA2D which maximizes the total covariance without distinguishing the inter-class covariance from the intra-class covariance, the groupings of images into subspaces would be performed not as a function of identity, but of facial expression, this leading to poor separation of the classes.
The LDA2D, jointly performing the minimization of the intra-class variance and the maximization of the inter-class variance, is therefore more robust to changes of facial expressions than the PCA2D.
Represented in
Thus, the bidimensional LDA, offering satisfactory classification by a reduction of dimensions, allies the discriminating power of the classical LDA and the saving in calculations of the PCA2D. Furthermore, with respect to these two procedures, the LDA2D is more robust to changes of pose and of facial expressions which are two major issues in face recognition.
Represented in
It is indicated in particular that the memory MEM of the device is devised to store at least:
As a variant, the device comprises a reader LEC of a removable memory medium (CD-ROM, ZIP, USB stick, or the like) able to store the program product and/or the learning base.
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/FR04/01395 | 6/4/2004 | WO | 12/4/2006 |