WO2006010855A1 - Method and device of face signature and recognition based on wavelet transforms - Google Patents

Abstract

The invention concerns a signature method of a face contained in a digital image (I) comprising: a step (E30) of structural representation of the face through an assembly of points; steps (E40, E50) of decomposition of the image into non-oriented and 2D-oriented wavelets; a step (E60) of determining the statistical characteristics of the image (I) obtained from coefficients of said wavelets in the neighbourhood of said points; and a step (E70) of signature of the face from said statistical characteristics.

Description

 Method and device for signing and recognizing a face based on wavelet transformations

The present invention is in the field of face recognition.

More specifically, the present invention relates to a method of signing a face represented by a digital image and a method of recognizing that face in a knowledge base from said signature. The methods according to the invention are based on transformation techniques of the aforementioned digital image in wavelets, in order to exploit the known properties of time-frequency decomposition.

A known method of face recognition based on wavelet transformation has been published in the article "Wavelet packet analysis for face recognition" on pages 289-297 of the journal "Images and Vision Computing, Vol. 18, 2000 ".

This method relies mainly on the calculation of the moments associated with three parts obtained by preliminary partition of the face (namely, the upper half of the center of the face, the lower part of the center of the face and the border of the face).

This method, which requires a frontal view of the face, has a major disadvantage in that it deals with relatively coarse blocks of the face, and does not exploit the local characteristics.

The invention aims to overcome the aforementioned drawbacks. For this purpose, and according to a first aspect, it concerns a process for signing a face contained in a digital image, this method comprising:

a step of structural representation of the face by a set of points; a step of decomposing the image into undirected wavelets and oriented 2D wavelets; a step of determining statistical characteristics of the image obtained from the coefficients of these wavelets in the vicinity of the aforementioned points; and

- a signature step of the face from these statistical characteristics.

The invention thus makes it possible to obtain a signature taking into account the characteristics of the face in the vicinity of each of the points of its structural representation.

Advantageously, the structural representation of the image in a set of points makes it possible on the one hand to limit the size of the signature of the face and, on the other hand, to take into account characteristics such as the luminance of the neighboring pixels.

Above all, the decomposition of the image into undirected wavelets and oriented 2D wavelets makes it possible to take into account, in the signature of the face, numerous details around the aforementioned points, and in particular details relating to the horizontal, vertical, diagonal contours to 45 °, 135 ° diagonals and their texture.

These characteristics make it possible to obtain a signature whose discriminating character is greatly improved with respect to all the techniques known from the prior art.

In a variant, it is also possible to use second generation wavelets ("2G wavelets").

The structural representation can include using a regular grid, or a set of randomly selected points. Preferably, the representation is based on a deformable model dedicated to the face object, which allows a representation of the face from a set of very significant points.

Preferentially, the wavelet decomposition uses a Daubechies base and a minimum resolution level equal to one sixteenth of the resolution of the initial image. Such a choice makes it possible to obtain a very satisfactory face recognition rate.

In a preferred embodiment, the statistical characteristics of the face comprise the mean and the standard deviation of the normalized coefficients of the wavelet associated with a mask around the points of structural representation.

As a variant, the statistical characteristics may comprise the sum of the standardized coefficients of the wavelets, or any higher order moment. In a preferred embodiment, the signature method according to the invention comprises a preliminary step of filtering the image, preferably with a Gaussian filter of size 5 out of 5.

This characteristic advantageously makes it possible to eliminate any disturbance caused by noise. According to a second aspect, the invention aims at a knowledge base comprising at least one signature of a face represented by a digital image, this signature comprising statistical characteristics of the image, obtained from the undirected wavelet decomposition coefficients. and 2D oriented of this image, in the vicinity of points of a structural representation of the face.

According to a third aspect, the invention relates to a method for recognizing a face represented by a digital image.

During this process:

a signature of the face is obtained by a signature method as briefly described above;

a comparison signature associated with a face is obtained in a knowledge base as briefly described above;

a distance is calculated between the two signatures, preferably quadratic; and

- It is decided whether the face to be recognized corresponds to the face associated with said comparison signature, from said distance.

In a first embodiment which is particularly simple to implement, a quadratic distance is used and it is considered in the calculation of the quadratic distance that each characteristic has the same importance.

This embodiment is particularly suitable when the structural representation of the face is made from a regular grid or sparsely chosen points on the face. Advantageously, this recognition method further comprises a step of weighting the statistical characteristics of the images to favor the most discriminating statistical characteristics in the calculation of the quadratic distance. In a first embodiment, these most discriminating characteristics are predetermined.

But in a preferred embodiment, a principal component analysis of the statistical characteristics of the comparison signatures of the knowledge base is performed to determine the coefficients of this weighting.

It is thus certain that the highest weighting coefficients actually correspond to the most discriminating characteristics.

Advantageously, the signature method according to the invention uses invariant characteristics of a predetermined number of points representing the digital image.

Such a signature can therefore serve as a key for indexing the digital image.

Accordingly, and in another aspect, the invention also relates to a method of indexing a digital image using a signature of this image of the type obtained by a signature method as briefly described above.

According to another aspect, the invention also aims at:

a digital signal representative of a signature of a face represented by a digital image, this signature comprising statistical characteristics of the image, obtained from the coefficients of non-oriented and 2D oriented wavelet decompositions of this image, in the vicinity; points of a structural representation of said face;

a transmission method and a method of recording a digital signal as briefly described above; a computer program stored on an information medium, this program including instructions enabling the implementation of a signature method and / or a recognition method and / or a method indexing as briefly described above when this program is loaded and executed by a computer system;

a device for signature of a face contained in a digital image, characterized in that it comprises: means for structural representation of this face by a set of points;

means for decomposing the image into undirected wavelets and oriented 2D wavelets;

means for determining statistical characteristics of the image obtained from the coefficients of these wavelets in the vicinity of the points; and

facial signature means based on these statistical characteristics; and

a device for recognizing a face represented by a digital image, characterized in that it comprises:

means for obtaining a signature of the face by a signature device as briefly described above;

means for obtaining a comparison signature associated with a face in a knowledge base as briefly described above;

means for calculating a distance, preferably quadratic, between these signatures; and

means adapted to decide whether the face to be recognized corresponds to the face associated with the comparison signature, from this distance.

The particular advantages of the knowledge base, the digital signal recognition method, the transmission method, the recording method, the computer program, the signature device and the recognition device are identical to those of the method of the invention. signature described above, they will not be recalled here.

Other aspects and advantages of the present invention will emerge more clearly on reading the description of particular embodiments which follows, this description being given solely by way of nonlimiting example and with reference to the appended drawings in which:

FIG. 1 represents, in flowchart form, the main steps of a signature process of a face represented in the form of a digital image, according to the invention;

FIG. 2 represents a regular grid that can be used to perform a structural representation of a face, in a particular embodiment of the invention;

FIG. 3 represents a deformable model that can be used to perform a structural representation of a face, in another particular embodiment of the invention;

FIGS. 4a and 4b respectively represent a decomposition of a wavelet image at resolutions VΣ and VA;

FIG. 5 represents, in flowchart form, the main steps of a face recognition method according to the present invention in a preferred embodiment; and

- Figure 6 shows a computer system incorporating a signature device and a face recognition device according to the invention in a preferred embodiment.

FIG. 1 represents, in flowchart form, the main steps E10 to E70 of a signature method Sv of a face V represented in the form of a digital image I, in accordance with a preferred embodiment of the invention. . In the following description, we will assume that this digital image I is sampled on a discrete grid of n lines and m columns, each pixel of the image being associated with at least one component Kx, y) of a representation space color.

In the embodiment described here, the component Kx, y) is a luminance value between 0 and 255.

The digital image I can therefore be considered as a discrete two-dimensional function, noted: [O ₁ m] x [O ₁ n] - * [0, 255] (x, y) -> I (χ, y)

The method described here comprises a prior step E10 of face detection in the digital image I.

This prior detection step E10 is not part of the invention and will not be detailed here.

For further details, those skilled in the art will be able to refer to the document "Detecting Faces in Images: A Survey" by MS Yang, D. Kriegman and N.Ahuja, published in the journal IEEE Transactions on Pattern Analysis and Machine Intelligence, Flight. 24, No. 1, pp. 34-58, in January 2002.

Whatever the case may be, the signature method according to the invention is independent of this detection step E10.

The detection step E10 is followed by a step E20 for filtering the digital image I.

In the preferred embodiment described here, this step consists in applying, to the digital image I, the Gaussian filter FG of size 5 × 5, according to:

In known manner, this pretreatment eliminates any disturbance of the image I, and in particular any disturbance caused by a possible noise.

The filtering step E20 is followed by a step E30 of structural representation of the face detected in the image I.

In general, this step consists of representing the face by a set of M points pi, so as to limit the size of the signature of the face and to bear the signature on discriminating points of the face.

Different types of structural representations can be used by the signature method according to the invention.

It can in particular be carried out: - using a regular grid, as shown in Figure 2;

- using a deformable model dedicated to the face object, as shown in Figure 3; or

- by random selection of a set of points. Of course, the structural representation based on a deformable model dedicated to the object face, allows to retain a set of points much more significant.

In the preferred embodiment, a regular grid is used, retaining one in ten points in each direction. The step E30 of structural representation of the face is followed by a step E40 of transforming the image I into unoriented wavelets, in order to obtain a representation of the image I at different resolutions 2 ', where j is a relative number less than or equal to -1.

A decomposition of the image I into wavelets at the resolutions V2 and ^λ A is shown diagrammatically respectively in FIGS. 4a and 4b.

In the preferred embodiment described herein, the non-oreinted wavelet transformation utilizes the Daubechies wavelet basis and a minimum resolution level of 1/16 (i.e., 2 ^-4 ).

For more details on the basis of Daubechies, the person skilled in the art will refer to the article "Ten readings on Wavelet" by Ingrid Daubechies, published in 1992 in the journal SIAM.

In known manner, the representation of the image I at a level of resolution 2 '(j ≤ -1), is given by a coarse image Ac ₂ ^j I and three images of details DcVl, DcVl and Dc ³ Z ¹ I, these images being all of the same size 2 ^{k + i} x 2 ^{1 + i} . In this undirected wavelet decomposition:

the image DcV1 is an image of details representing the vertical high frequencies of the image I;

the image DcV1 is an image of details representing the horizontal high frequencies of the image I; and the DcV1 image is an image of details representing the diagonal high frequencies of the image I. The step E40 of transforming the image I into undirected wavelets is followed by a step E50 of transforming the image I into oriented 2D wavelets.

Such decomposition is described in Kim N. D and Udpa S.'s "Texture Classification Using Roated Wavelet Filters" published in 2000 in the journal IEEE Transactions on Systems, Mans and Cybernetics, Vol. 6, No. 6 at pages 847-852.

In the preferred embodiment described here, this oriented 2D wavelet transformation also uses the Daubechies wavelet basis and a minimum resolution level of 1/16.

In known manner, the 2D wavelet decomposition oriented of the image I at a resolution level 2 '(j ≤ -1), is given by a coarse image Ao ₂ ' I and three images of details DoVl, Do ² ₂ ' I and Do ³ ₂ 'I, these images being all of the same size 2 ^{k + j} x 2 ^{1 + j} . In this 2D oriented wavelet decomposition:

- The image Do ¹ ₂ 'I is a detail image representing the 45 diagonal high frequencies of the image I;

the image D ² ₂ ^j 1 is a detail image representing the diagonal high frequencies at 135 ° of the image I; and the image C ³² ₂ 'I is an image of details representing the other diagonal frequencies of the image I.

The step E50 for transforming the image I into oriented 2D wavelets is followed by a step E60 for determining the dominant characteristics of the image I. This step E60 consists in determining statistical characteristics that best represent the neighborhood of each of the Ms. points pi obtained during the step E30 structural representation of the face.

In the preferred embodiment described here, this step E60 comprises two sub-steps E62 and E64. The first sub-step E62 consists of determining statistical characteristics that best represent the neighborhood of each of the points pi obtained during the step E30 of structural representation of the face. During the first sub-step E62, for each wavelet decomposition (four non-oriented and four oriented 2D) obtained previously, a mask Q, ¹ is calculated around each of the points p, where 1 is the resolution level of the decomposition considered, between 1 and 4. In a preferred embodiment, this mask Q, ¹ is rectangular.

During the second substep E64, for each point Pi, and for each of the eight wavelet decompositions at resolution 1, the mean and the standard deviation of the wavelet coefficients associated with the mask Q, ¹ are calculated.

In this preferred embodiment, the means and standard deviations thus obtained are normalized by the maximum value of the coefficients of the subband processed.

These standard averages and standard deviations, respectively denoted m, ¹ and σ, ¹ , constitute the dominant characteristics of image I.

In the rest of the description, we will call s ,, signature of the point p ,, the set: s, = {s, ¹ }, where s, ¹ = (m, ^, σ ,,, ¹ )} with

1 = 0, ... 4, and j = Ac ₂ ¹ I, DcVl, Dc ² ₂ 'I; Dc ³ ₂ ^J I, Ao ₂ ¹ I, DoVl, Do ² ₂ ^J I, Do ³ ₂ I. The step E60 for determining the dominant characteristics is followed by a signature step E70 of the face contained in the image I. In the embodiment described here, this signature Sv is constituted by the set of signatures s, points p , obtained in the previous step:

Sv = {s ,, i = 1 ... M}

Signature step E70 completes the face signature method described herein. This signature Sv can in particular be used in a face recognition method V according to the invention.

We will now describe, with reference to FIG. 5, the main steps E100 to E 160 of a face recognition method V according to the present invention in a first preferred embodiment. More specifically, this recognition method makes it possible to verify the existence of the face V in a DB knowledge base according to the present invention. According to the invention, the knowledge base DB comprises a set of K signatures S _v ^k of faces V ^k , these signatures being of the type of the signature Sv described above with reference to FIG. 4.

Sv ^k = {s _i \ i = 1 ... lv1} ₎ k = 1 ... K In a first preferred embodiment, the recognition method according to the invention identifies the face V as being identical to a face V ^k of the knowledge base DB, when the quadratic distance d between the signatures S _v and Sv ^k of the faces V and V ^k is less than a first predetermined threshold Th 1. In a classical way, the quadratic distance d expresses as follows:

MM d = Σ2> / - * / *) ² . or :

- M is the number of points in the structural representation of the face;

- | S | denotes the size of the signature vector Sj associated with each point Pi of the structural representation;

- s / denotes the j ^èmθ component of the signature of the i ^θmθ point of the structural representation of the face V; and

- If ^ik denotes the j ^th component of the signature of the ^i-th point of the structural representation of the face V ^k. In known manner, the distance between the two signature vectors si and Si ' ^{k is} expressed directly from the sum of the squares of the differences of the components mg ¹ , σ _{i j} ¹ of these vectors.

In a first variant embodiment, the threshold Th 1 is chosen empirically. In this first variant, the distribution of the distances d is studied and the threshold Th 1 is fixed such that if the distance between two faces is greater than Th1, then the faces are considered dissimilar.

In a second variant embodiment, the threshold Th 1 is chosen after a statistical analysis of the distribution of distances d.

In a preferred embodiment of this second embodiment, a Bayesian analysis of this distribution of distances is carried out. Such a method, known to those skilled in the art, consists in studying the for a set of learning constituting a knowledge a priori, the distribution of distances d for faces, of the knowledge base DB wrongly recognized, rightly recognized , unrecognized and not properly recognized.

We will now describe, with reference to Figure 5, a method of recognition of a face V, according to a second preferred embodiment of the invention.

Here again, this recognition process aims to identify whether the face V is identical to a face V ^k of the DB knowledge base.

In the example described here, it will be assumed that the knowledge base DB has the signatures Sv ^k of K faces.

This second preferred embodiment is based on a principal component analysis, such analysis being known in the field of statistics.

This recognition method comprises a first step E100 for creating an X matrix comprising the signatures Sv ^k .

More precisely, the matrix X is written:

X = s / ... _S1 ^K

The step E100 for creating the matrix X is followed by a step E110 for calculating the covariance matrix Σ of the matrix X.

In known manner, the covariance matrix Σ is obtained as follows:

Σ = (X - μ) (X - μ) ^τ , where μi = -X s}

P H

The step E110 for calculating the covariance matrix Σ is followed by a step E120 for determining the eigenvalues and eigenvectors of the covariance matrix Σ. The calculation of eigenvalues and eigenvectors of the matrix S is known and will not be detailed here.

Note U the column matrix of K eigenvectors.

Note respectively SV and SV ^k the projection of the signature Sv and Sv ^k in the eigenvector space.

In known manner, S ' _v = U. Sv; and S ' _v ^k = U. S _v ^k

The step E120 for determining the values and the eigenvectors is followed by a step E130 of normalization of the eigenvalues. This step makes it possible to obtain M coefficients λi, ... λ _m , between 0 and 1. The normalization step E130 is followed by a loop during which a weighted quadratic distance dp is calculated between the signatures of the faces V and V _k .

This loop comprises a step E 140 of the calculation, strictly speaking, of this weighted quadratic distance dp, with:

- s? denotes the j ^th component of the projection of the signature of the i ^th point of the structural representation of the face in the space V of eigenvectors; and

if ^{i ij} designates the j ^th projection component of the signature of the i ^th point of the structural representation of the face V ^k in the space of the eigenvectors.

Thus, according to the known characteristics of the principal component analysis, the weighted quadratic distance dp makes it possible to attach more importance to the most discriminating statistical characteristics. Step E 140 is followed by a test E 150 in which the weighted quadratic distance dp is compared with a second predetermined threshold Th2.

The choice of the Th2 threshold may in particular be made similarly to the choice of Th threshold Th 1 described above.

If, during this test, it is determined that the quadratic distance dp is less than or equal to the second threshold Th2, it is deduced, during a step E 160, that the face V is recognized in that it comprises a signature similar to that of the face V ^k . In the opposite case, we go back to the previous step E140 to compare the signatures of the faces V and V ^{k + 1} .

Of course, the recognition process ends when the signature of the face V has been compared without success with the signatures of the faces K ^k .

FIG. 6 represents a computer system 100 incorporating a signature device and a device for recognizing a face V according to the invention.

The computer system 100 described here consists of a personal computer of the PC type.

This computer system 100 comprises, in a known manner, a processor 10, a screen 11, a network card 12, a read-only memory ROM 13, a random access memory RAM 14, these various elements being, in a known manner, connected by In the embodiment described here, the computer system 100 is connected, by means of the network card 12, to a computer network 20 to which is connected a knowledge base DB according to the invention.

This knowledge base DB comprises a plurality of signatures S _v ^k of faces V ^k of the type obtained by the signature method described above, with reference to FIG.

More specifically, each signature Sv ^k of the knowledge base DB has statistical characteristics (my ¹ , σ- ^) of an image V ^k obtained from the classical and 2D wavelet decomposition coefficients of a digital image representing the face V ^k in the vicinity of points pi of a structural representation of this face.

In the embodiment described here, the computer system 100 is connected to a digital camera 16 adapted, in cooperation with the other elements of the computer system 100, to acquire a digital image I and to store this image in a register of the same name. RAM 14. The ROM 13 includes registers storing a computer program PG 1 adapted to implement the signature method described above with reference to Figure 1. This computer program PG 1 is thus adapted to detect, if necessary, a face V in the digital image I stored in the RAM memory

14, to filter this image using the FG filter stored in the read-only memory

13 and to obtain a structural representation by a set of M points pi.

The signature program PG 1 is also adapted to carry out an undirected wavelet decomposition and an oriented D wavelet and to determine the dominant characteristics of the image I as described above, with reference to the step E60 of FIG. The signature method PG 1 is thus adapted to obtain the signature Sv of the image V and to store this signature in a register of the same name of the random access memory 14.

The read-only memory 13 of the computer system 100 also comprises a program PG2 adapted to create a signal SG of digital data representative of the signature Sv and to transmit, in collaboration with the network card 12, this signal SG to another computer system not shown here by means of a communication protocol, preferably standard.

The computer system 100 also includes a PG3 program for recording the SG signal. In the preferred embodiment described here, the recording program PG3 is able to record the signal SG in the knowledge base DB on the computer network 20.

The ROM 13 of the computer system 100 also comprises a computer program PG4 for recognizing a face V represented by a digital image I.

This recording program PG4 is adapted to implement a recognition method as described above with reference to FIG.

For this purpose, the recognition program PG4 is adapted to obtain signatures Sv ^k of face V ^k from the knowledge base DB, via the network card 12 and the computer network 20 and to store these signatures in the random access memory 14 . The recognition program PG4 is in particular adapted to carry out a principal component analysis of the statistical characteristics of the signatures S _v ^k of the knowledge base DB so as to obtain the weighting coefficients λj described above, with reference to FIG. The recognition PG4 is then adapted to calculate, from these weighting coefficients, a weighted quadratic distance between the signature Sv of the face V and the different signatures Sv ^k obtained from the knowledge base DB.

In the preferred embodiment described here, the recognition program PG4 is adapted to display on the screen 11 a predetermined message when the aforementioned quadratic distance is less than a threshold Th1 or Th2 stored in the read-only memory 13, this message being significant because the face V is recognized in that it has a signature similar to that of the face V _k associated with the signature Sv ^k -

Claims

1. A method of signature of a face (V) contained in a digital image (I), characterized in that it comprises: a step (E30) of structural representation of the face (V) by a set of points (p,); a step of decomposing the image (I) into non-oriented wavelets (E40) and oriented 2D wavelets (E50); a step (E60) for determining statistical characteristics (m ,,, ¹ , α ,,, ¹ ) of the image (I) obtained from the coefficients of said wavelets in the vicinity of the points (p,); and a step (E70) of signature (Sv) of the face (V) from said statistical characteristics (m ,, /, α ,,, ¹ ). 2. Signature method according to claim 1 characterized in that said structural representation step (E30) uses: - a regular grid; or - a deformable model dedicated to the face object; or - a random selection of points. 3. Signature method according to claim 1 or 2 characterized in that said decomposition (E40, E50) uses a Daubechies base and a minimum resolution level equal to one sixteenth of the resolution of said image (I). 4. Signature method according to any one of claims 1 to 3, characterized in that said statistical characteristics (m ,,, ¹ , σ ,,, ¹ ) comprise the mean (m ,,, ¹ ) and the difference deviation _(σ,, /) of normalized coefficients of said wavelet associated with a mask (Q ¹⁾ around said points (p,). 5. Signature method according to any one of claims 1 to 4, characterized in that it comprises a preliminary step (E20) filtering said image (I), preferably with a Gaussian filter size 5 of 5. 6. A method for recognizing a face (V) represented by a digital image (I), characterized in that a signature (Sv) of said face (V) is obtained by a signature method according to any one of claims 1 to 5; a comparison signature (Sv ^k ) associated with a face (V ^k ) is obtained in a knowledge base (DB); calculating (E140) a distance (d, dp), preferably quadratic, between said signature (S _v ) and said comparison signature (Sv ^k ); and - it is decided (E160), if said face (V) corresponds to the face (V ^k ) associated with said comparison signature (Sv ^k ), from said distance (d, dp). 7. A method of recognition according to claim 6 wherein a weighted quadratic distance is used and characterized in that it further comprises a step (E 140) for weighting the statistical characteristics (rrij _j ¹ , σ _{i j} ¹ ) of said images for to favor the most discriminating statistical characteristics in the calculation of said weighted quadratic distance (dp). 8. Recognition method according to claim 7, characterized in that a principal component analysis of said statistical characteristics (mi, /, σj _j ¹ ) of the comparison signatures (Sy ^k ) of the knowledge base (DB ) to determine the coefficients (λι) of said weighting. 9. Recognition method according to any one of claims 6 to 8, characterized in that, during said decision step, a decision threshold (Th1, Th2), obtained from a statistical analysis, is used. preferentially bayesian, of the distribution of a plurality of distances (d, dp) between faces of the knowledge base (DB). 10. A method of indexing a digital image (I), characterized in that said indexing uses a signature of said digital image (I) of the type obtained by a signature method according to any one of claims 1 to 5. 11. A digital signal representative of a signature (Sv ^k ) of a face (V ^k ) represented by a digital image, said signature being obtained by a signature method according to any one of claims 1 to 5. 12. A method of transmitting a digital signal (SG) according to claim 11. 13. A method of recording a digital signal (SG) according to claim 11. 14. Computer program stored on an information medium (13), said program comprising instructions for implementing a signature method according to any one of claims 1 to 5 and / or a method recognition system according to any one of claims 6 to 9 and / or an indexing method according to claim 10 when this program is loaded and executed by a computer system (100). 15. Device for signature of a face (V) contained in a digital image (I), characterized in that it comprises: means (10, 13, 14) for structural representation of the face (V) by a set of points (pi); means (10, 13, 14) for decomposing the image (I) into non-oriented wavelets and oriented 2D wavelets; means (10, 13, 14) for determining statistical characteristics (rrijj ¹ , σy ¹ ) of the image (I) obtained from the coefficients of said wavelets in the vicinity of the points (pi); and means (10, 13, 14) for signature (Sv) of the face (V) from said statistical characteristics (rrijj ¹ , σi _j ¹ ) - 16. Device for recognizing a face (V) represented by a digital image (I), characterized in that it comprises: - means (10, 13, 14) for obtaining a signature (S _v ) said face (V) by a signature device according to claim 15; means (10, 13, 14) for obtaining a comparison signature (Sv ^k ) associated with a face (V ^k ) in a knowledge base (DB); means (10, 13, 14) for calculating a distance (d, dp), preferably quadratic, between said signature (Sv) and said comparison signature (Sv ^k ); and means (10, 13, 14) adapted to decide whether said face (V) corresponds to the face (V ^k ) associated with said comparison signature (Sv ^k ), from said distance (d, dp).