WO2012175391A1

WO2012175391A1 - Method of establishing a final score of similarity between images

Info

Publication number: WO2012175391A1
Application number: PCT/EP2012/061227
Authority: WO
Inventors: Liming Chen; Di HUANG; Mohsen Ardabilian; Yunhong Wang
Original assignee: Ecole Centrale De Lyon; Institut National Des Sciences Appliquées De Lyon; Université Lyon 1 Claude Bernard; Université Lyon 2 A. Et L. Lumière; Centre National De La Recherche Scientifique; Beijing University Of Aeronautics And Aerospace
Priority date: 2011-06-22
Filing date: 2012-06-13
Publication date: 2012-12-27
Also published as: FR2977056B1; FR2977056A1

Abstract

This method of establishing a final score of similarity between a digital image acquired and a pre-recorded digital image, comprises: a) the construction (42) on the basis of the image acquired of H maps J_i, of gradients calculated according to a respective set i of directions, b) for each map of gradients J_i, the calculation (52) of an intermediate score S_i, of similarity between this map of gradients J_i, and a corresponding reference map J_iref, each intermediate score S_i of similarity being a real number representing the degree of similitude between the compared maps, the reference map J_iref being a map of gradients calculated according to the same set i of directions as the map of gradients J_i, on the basis of the pre-recorded image, and c) the establishment (54) of a final score S_f of similarity between the image acquired and the pre-recorded image as a function of the H intermediate scores S_i of similarity.

Description

METHOD OF ESTABLISHING A FINAL SIMILARITY SCORE BETWEEN

IMAGES

[Ooi] The invention relates to a method for establishing a final similarity score between a digital image of an object acquired by an electronic sensor and a prerecorded digital image in an image library. The invention also relates to a recording medium for implementing this method as well as to a method and an apparatus for recognizing an object.

[002] A digital image is formed of pixels P (x, y) and comprises a matrix which at each position (x, y) of a pixel in the image associates at least one value a (x, y) of a physical quantity acquired or measured by the electronic sensor.

[003] A pixel is the smallest homogeneous surface of a 2D or 3D image.

Generally, in an image, pixels are aligned in rows and columns. Subsequently, the direction in which the lines extend is called horizontal while the direction in which the columns extend is said to be vertical.

[004] In a 2D image, the position of each pixel is encoded by two coordinates in a two-dimensional coordinate system.

[005] In a 3D image, the position of each pixel is coded by three coordinates in a three-dimensional coordinate system. This allows to code more depth.

[006] A texture image, better known by the term "image texture", is an image in which the value a (x, y) is a value encoding a color or the power of a radiation in a predetermined range of radiation. This value a (x, y) can be acquired by different types of sensor. For example, the sensor may be an infrared sensor. It can also be a light intensity or a color coded by a triplet of information.

[007] A range telemetric image, better known by the term "range image", associates at each position of a pixel a value a (x, y) which represents the distance that separates the point from the corresponding object. at this pixel of a reference plane such as that attached to the lens of the electronic sensor.

[008] In the remainder of this description, when it is necessary to distinguish a data item relating to a texture image from that relating to a range telemetric image, this datum includes the index "T". Conversely, a datum relative to a telemetric depth image includes the index "R". For example, the value a _T (x, y) corresponds to the value a (x, y) in the texture image and the value a _R (x, y) corresponds to the value a (x, y) in the telemetric depth image. In the case where the data has neither the index "T" nor the index "R", it means that it concerns either a texture image or a telemetric image.

[009] Known methods for establishing a final similarity score include: a) constructing from the acquired image of H maps J, gradients calculated according to a respective set of directions, where:

H is an integer, of a set i of directions, greater than or equal to two,

i is an identifier of the set of directions in which the gradients are calculated, a set of directions associates with each pixel P (x, y) of the image a direction Oi (x, y), the direction Oi (x , y) being angularly shifted relative to the direction Oi (x, y) by an angle θ, this angle Θ, being the same regardless of the pixel considered in the image,

a map of gradients calculated according to a set of directions being a matrix which at each pixel P (x, y) of the acquired image associates a value ρ, (χ, ν) representative of the gradient of the physical quantity in the direction Ο, (χ, ν) associated with this pixel in the game i.

[ooi o] A map of gradients calculated according to a direction i is a matrix which at each pixel P (x, y) of the acquired image associates a value ρ, (χ, ν) representative of the gradient of the physical quantity in the direction i from this pixel.

An example of a known method is for example disclosed in US patent application 2010/0135540.

The methods for establishing a final similarity score between two images are particularly useful for comparing two images and for recognizing an object in an image.

In the known methods, a matrix of gradients between an image I and reference images is constructed in the horizontal and vertical direction. Then a final similarity score is calculated directly from this difference matrix.

[0014] State of the art is also known from:

- Di HUANG and Al: "A novel geometry facial representation based on multi-scale extended local binary patterns", Automatic Face & Gesture Recognition and Workshops (FG 201 1), 201 1, IEEE, March 21, 201 1, pages 1 -7 , and

- WAEL BEN SOLTANA et al: "Adaptive Feature and Score Level Fusion Strategy Using Genetic Algorithms," Pattern Recognition (ICPR), 2010 20th International Conference on, IEEE, Piscataway, NJ, USA, August 23, 2010, pages 4316-4319.

These known methods work properly. However, it is necessary to improve the accuracy and precision of these final scoring methods, for example, to achieve recognition rates of an object in an even higher image.

The invention therefore aims at providing an improved method for establishing a final similarity score in which the method comprises:

b) for each gradient map J ,, the calculation of an intermediate score S, of similarity between this gradient map J, and a corresponding reference map J _ire f, each intermediate score S, of similarity being a real number representing the degree of similarity between the compared maps, the reference map J _iref being a map of gradients calculated according to the same set of directions as the gradient map J, from the prerecorded image, and

c) establishing a final score S _f of similarity between the acquired image and the prerecorded image according to the H intermediate scores S, of similarity.

Calculating intermediate similarity scores between gradient maps calculated in different directions increases the accuracy and accuracy of the final similarity score. Therefore, a method of recognizing an object in an image using the final scores established according to the above method obtains a higher exact recognition rate.

In addition, since gradient maps are used to establish the final similarity score, this method is robust to changes, for example during the acquisition of the image, which adds a constant to the value a (x, y). For example, such a change may be a more distant shot.

Finally, the above method is generic. Indeed, it can be applied to any type of image such as an image of a face, a fingerprint, an iris of an eye, veins, a landscape and other .

Embodiments of this method for establishing a final similarity score may include one or more of the following features:

The construction of the map J, of gradients comprises the normalization of the gradient associated with each pixel of this map by dividing it by at least one component or a combination of components of a vector of directional gradients, this component or this combination of components being such that if each component of the vector of directional gradients and this gradient are multiplied by the same constant λ, the normalized gradient remains independent of this constant λ, the vector of directional gradients gathering only gradients values, in different directions, associated at that same pixel;

■ the construction of the map J, of gradients comprises the normalization of the gradient associated with each pixel of this map by dividing it by the norm of the vector of directional gradients or the maximum or the minimum or a linear combination of the components of this same vector directional gradients;

The method comprises determining the direction Oi (x, y) associated with each pixel P (x, y) in the set of directions, from the gradients, in at least two different directions, of the neighboring pixels included in a predetermined perimeter centered on this pixel P (x, y);

The direction Oi (x, y) associated with each pixel P (x, y) in the set of directions is the same regardless of the pixel of the image; The construction of each map J, of gradients comprises the smoothing of the gradient associated with each pixel, said central pixel, of this map by realizing a weighted average of the gradient of the central pixel with those of the adjacent pixels situated inside a predetermined perimeter centered on the central pixel;

Weighting coefficients are assigned to each gradient of one pixel of the weighted average, the weighting coefficient assigned to a gradient of an adjacent pixel being smaller when the adjacent pixel is remote from the central pixel;

■ the steps a) and b) are carried out on a texture image and on a telemetric depth image of the object and, in step c), the final score S _f is established from the intermediate scores S, obtained from the texture image and the depth telemetric image;

■ the establishment of the final score S _f comprises the realization of a weighted sum of intermediate scores S ,, a different weight being assigned to each intermediate score of this sum;

The weights of the weighted sum are selected by implementing a genetic algorithm;

■ H is an integer greater than or equal to two if the gradient maps have negative and positive values or four if the gradient maps have only null or positive values.

Embodiments of this method for establishing a final similarity score furthermore have the following advantages:

the normalization of the vector of directional gradients makes the process robust vis-à-vis the changes which, for example during the acquisition of the image, introduce a multiplying coefficient on the value a (x, y) of the physical quantity acquired;

the use of sets of directions constructed from the gradients of the neighboring pixels makes it possible to make the method robust with respect to a rotation of a local group of pixels of the image around an axis perpendicular to the plane of the image ;

- The smoothing of directional gradients removes abrupt changes of directional gradients of a pixel to its neighbors, created, in particular, by the scanning of the image of the object;

during smoothing, using weighting coefficients which are all the smaller as the corresponding adjacent pixel is removed from the central pixel makes it possible to increase the robustness of the method;

the use of the Gaussian law to assign a value to these weighting coefficients makes it possible to increase the accuracy of the final score of similarity and thus the exact recognition rate of an object when this method is used for this purpose, the combined use of a texture image and a depth telemetric image of the same object improves the recognition rate of this object when the final similarity score is used for this purpose.

The invention also relates to a method for recognizing an object, this method comprising:

the acquisition by an electronic sensor of a digital image of the object to be recognized, each digital image being formed of pixels P (x, y) and comprising a matrix which associates at each position (x, y) with a pixel P (x, y) in the image at least one value a (x, y) of a physical quantity measured by the electronic sensor,

calculating a final score S _f of similarity between the acquired image and each prerecorded image of a digital image library by implementing the above method of establishing a final similarity score, and

- Establishing the correspondence between the acquired image and one of the prerecorded images of the library based on the final score S _f similarity calculated between these two images, the object being recognized when the acquired image of this object corresponds to one of the pre-recorded images.

The invention also relates to an information recording medium comprising instructions for the implementation of the method above, when these instructions are executed by an electronic computer.

Finally, the object of the invention is a device for recognizing an object, this device comprising:

an electronic sensor capable of acquiring a digital image of the object to be recognized, each digital image being formed of pixels P (x, y) and comprising a matrix which associates with each position (x, y) of a pixel in the image of at least one value a (x, y) of a physical quantity measured by the electronic sensor,

an electronic calculator programmed for:

Construct, from the acquired image, H maps J, gradients calculated according to a respective set of directions, where:

H is an integer, of games i of directions, greater than or equal to two,

i is an identifier of the set of directions in which the gradients are calculated,

a set of directions associates with each pixel P (x, y) of the image a direction Oi (x, y), the direction Oi (x, y) being angularly offset with respect to the direction Oi (x, y) ) an angle θ ,, this angle Θ, being the same whatever the pixel considered in the image,

a map of gradients calculated according to a set of directions being a matrix which at each pixel P (x, y) of the acquired image associates a value ρ, (χ, ν) representative of the gradient of the physical quantity in the direction Ο, (χ, ν) associated with this pixel in the set i,

• a) for each gradient map J ,, calculate an intermediate score S, of similarity between this map of gradients J, and a corresponding map J _iref , each intermediate score S, of similarity being a real number representing the degree of similarity between the compared maps, the reference map J _ire f being a map of gradients calculated according to the same set of directions as the gradient map J, from a prerecorded image of a digital image library, and

· B) establishing a final score S _f of similarity between the acquired image and the prerecorded image as a function of the H intermediate scores S, of similarity,

Repeating the steps a) and b) for each prerecorded image of the library and establishing the correspondence between the acquired image and one of the prerecorded images of the library according to the final score S _f of similarity calculated between these two images, the object being recognized when the acquired image of this object corresponds to one of the prerecorded images.

The invention will be better understood on reading the description which follows, given solely by way of nonlimiting example and with reference to the drawings in which:

FIG. 1 is a schematic illustration of a biometric access control system;

Figure 2 is an illustration of different directions used in the system of Figure 1;

FIG. 3 is a flowchart of an access control method using the system of FIG. 1;

FIG. 4 is a schematic illustration of the Gaussian law used in the method of FIG. 3.

In these figures, the same references are used to designate the same elements.

In the following description, the features and functions well known to those skilled in the art are not described in detail.

FIG. 1 represents a biometric system 2 for controlling access to a resource 4. By way of illustration, the resource 4 is a building.

The system 2 is described in the particular case where, for a given index i, all directions Oi (x, y) are parallel. We denote O, this common direction.

Access to this resource is authorized by the system 2 only if the user is an authorized user. An authorized user is a user whose biometric characteristics are associated, by a database, with a right to use this resource 4. Here, the biometric features are the face of the user.

For this purpose, the system 2 comprises a device 6 for recognizing an object by comparing a digital image of this object to a library of prerecorded digital images. In this particular case, the object is a face. The device 6 comprises an apparatus 8 for acquiring digital images of the user's face. In this particular case, the apparatus 8 comprises an electronic sensor 10 able to acquire a texture image of the user's face. For example, the sensor 10 is an infrared camera that measures the intensity of the infrared radiation at a multitude of points on the user's face.

The apparatus 8 also comprises another electronic sensor 12 for acquiring a range telemetric image of the user's face. This sensor 12 is typically a range finder that measures the distance that separates points of the user's face from a reference plane. For example, the sensor 12 uses laser telemetry technology. Typically, tens of thousands of facial points are taken into account to acquire the telemetric depth image.

The sensors 10 and 12 are connected to a central unit 14. The central unit 14 is programmed to allow access to the resource 4 if the image of the user acquired by the sensors 10 and 12 corresponds to that an authorized user. In the opposite case, the central unit 14 prohibits access to the resource 4.

The central unit 14 is made from a programmable electronic computer 16 capable of executing instructions recorded on a recording medium. For this purpose, the computer 16 is connected to a memory 18 containing the instructions necessary to execute the method of FIG.

The memory 18 also contains the database 20 which associates the biometric characteristics of the users with a right to use the resource 4. Here, it is assumed that the authorization to access the resource 4 is given as soon as the The user's biometric characteristics are recorded in the database 20. Thus, this database 20 does not need to encode the right to use the resource 4 in a particular field. But one can imagine applications where the authorization to access a resource varies according to recognized people. It is then necessary to code the corresponding right of use.

The biometric characteristics recorded in the database 20 for each authorized user may be digital images of the user's face, vein images, face range (range image) or any combination of those images. -this. For example, for each authorized user, a texture image and a telemetric image of depth of his face can be recorded in the base 20. In this case, the base 20 contains:

for each texture image, H reference maps J _ire fT calculated from this texture image, and

- for each range image, H reference maps J _ire fR calculated from the range image.

H is an integer number of sets of directions greater than or equal to two and, preferably, greater than or equal to four, eight or sixteen. In the rest of this description, H is taken equal to eight. We denote N the number of cards _Jire f associated with the same authorized user. Here this number N is equal to 2Ή designating the number of maps calculated on the texture image and the telemetric depth image.

Each reference map J _ire f is a map of gradients calculated according to a respective direction i. Here, eight different directions are used. These eight directions are represented in Figure 2 by arrows Oi O _8. Here, these directions Oi O ₈ are uniformly distributed in the image plane. In other words, the angle Θ between two successive directions is constant. Here this angle Θ is equal to 45 °. The direction Oi is the horizontal direction.

FIG. 2 also represents a square pixel P (x, y) whose geometric center is marked by a point O. In the horizontal direction, this pixel P (x, y) is followed by the pixel P (x + 1). , y) and preceded by the pixel P (x-1, y). In the vertical direction, this pixel P (x, y) is followed by the pixel P (x, y + 1) and preceded by the pixel P (x, y-1). In the diagonal direction O ₂ , the pixel P (x, y) is followed by the pixel P (x + 1, y + 1) and preceded by the pixel P (x-1, y-1). In the diagonal direction O ₄ , the pixel P (x, y) is followed by the pixel P (x-1, y + 1) and preceded by the pixel P (x + 1, y-1).

The calculation of the reference maps J _iref is described in more detail with reference to FIG. 3.

The operation of the system 2 will now be described with reference to the method of FIG. 3.

Initially, during a step 40, the user 40 presents his face in front of the sensors 10 and 12. In response, the sensors 10 and 12 acquire a digital texture image and a telemetric digital depth image of this face . For this, these sensors measure the values a (x, y) for each pixel of each image. The acquired images are recorded for processing in the memory 18. In the texture image, _T (x, y) is a function of the light intensity received by the sensor. In the range telemetric image, a _R (x, y) is a distance measured by the sensor 12. The combination of the texture image and the telemetric image forms a textured 3D image in which the position of each pixel is coded by a triplet (x, y, z) and at each position the value a _T (x, y) is associated. The z coordinate is set from the value a _R (x, y). In a step 42, the computer 16 builds H cards J _iT gradients along the H directions of Figure 2 from the acquired texture image.

More specifically, during an operation 44, the computer 16 constructs for the acquired texture image H gradient maps _T to L _HT . Each card L _iT contains for each pixel P (x, y), the gradient or the derivatives δα _τ (χ, ν) / δί of the value a _T (x, y) in the direction i, where i is the identifier from the direction that the gradient map is calculated. For example, each map L _iT associates with the position (x, y) of a pixel P (x, y) the value of the gradient ρ, τ (χ, γ) starting from this pixel to the pixel immediately consecutive in the direction i.

Here, all the directions taken into account, includes for each direction i its opposite direction. Under these conditions in each L _iT card, only the positive values of each gradient are retained. When the gradient associated with a pixel P (x, y) is negative, it is replaced by a null value.

For example, in the case of the positive horizontal direction Οι, Ριτ (χ, γ) is calculated using the following relation: pn- (x, y) = (a _T (x + 1, y ) - _T (x-1, y)) / 2. Preferably, the same relationship is used in all directions.

Then, during an operation 46, from each L _iT card, a new gradient map L _iT ^R is built. In the L _iT ^R map the abrupt variations of gradients ρ, τ (χ, γ) between two neighboring pixels are smoothed.

For example, for this, the computer 16 achieves the weighted average of the value pn- (x, y) with the gradients of the pixels contained in a predetermined perimeter centered on the center O of the pixel P (x, y).

For example, the predetermined perimeter is a circle of radius R. In addition, the weighting coefficient assigned to each gradient is even smaller than the corresponding pixel is away from the pixel P (x, y). Preferably, the weighting coefficients follow a Gaussian law.

This average value of the gradient pn- (x, y) is here denoted p _iT ^R (x, y).

By way of illustration, it is assumed that the radius R chosen is such that the perimeter encompasses only the pixels represented in FIG. 2. In this case, the gradient p _iT ^R (x, y) is calculated according to the relation next :

p _iT ^R (x, y) = cipi (x, y) + c ₂ pi (x + 1, y) + c ₃ pi (x, y + 1) + c ₄ pi (x-1, y) + c ₅ μl (x, y-1) + c ₆ μl (x + 1, y + 1) + c ₇ μl (x-1, y + 1) + c ₈ μl (x-1, y-1) + c ₉ ft (x + 1, y-1)

The coefficients c, are chosen according to the Gaussian law shown in FIG. 3. In this figure d is the distance between the center O of the pixel P (x, y) and the centers of the neighboring pixels. The Gaussian law is symmetrical. For the sake of clarity, the coefficients c, have been represented either to the right or to the left of this ordinate axis. The chosen Gaussian law is parameterized by the radius R of the perimeter.

From a practical point of view, a Gaussian core G whose standard deviation is proportional to the radius R is used to calculate the values p _iT ^R (x, y). By For example, the gradient map L _iT is convolved with this core G to obtain all the values p _iT ^R (x, y) of the map L _iT ^R.

For example, the values of the core G are given by the following relation:,

Where:

- "exp" is the exponential function,

- x-Xc II is the distance between the center O and the adjacent pixel,

- "sigma" is the standard deviation chosen according to the distance R.

Subsequently, we denote by ^R (x, y) the vector of directional gradients [pi _T ^R (x, y),

PHT ^R (x, y)] ^T , where the exponent "T" is the transposed symbol.

Then, during an operation 48, each L _iT ^R card is normalized. The map

Standardized _iT ^R is rated J _iT . This J _iT card is called the perceived facial image.

Normalization consists in dividing the value p _iT ^R (x, y) of a pixel P (x, y) by the norm of the vector p _T ^R (x, y), associated with this same pixel P (x , y). The normalized value of the pixel P (x, y) in the card J is denoted p _iT ^N (x, y).

Here, the standard | p _T ^R (x, y) | of the vector pr ^R (x, y) is given by the classical definition, that is to say the square root of the sum of its components squared: | p _T ^R (x, y) | = sqrt (Σpi _T ^R (x, y) ² ) for i varying from 1 to H, where "sqrt" is the square root function.

Normalization makes it possible to make the above method robust with respect to any modification of the environment during the acquisition of the image which multiplies the value pn- (x, y) by a constant λ. . Indeed, the value p _iT ^N (x, y) becomes independent of this constant λ. Thus, this method is robust vis-à-vis changes such as a change in light contrast during step 40. Moreover, in this method, λ is not necessarily constant over the entire image and may vary. from one pixel to another. Thus, the method is insensitive to local changes such as a local change of contrast during the acquisition of the image.

The step of constructing the gradients map ends when the gradient map J _iT has been constructed for each of the H directions.

After step 42, during a step 52, the calculator 16 calculates intermediate scores S _iT of similarity between each J _iT card and the corresponding _Jir fT card.

[0064] The J f _ire cards are constructed as described in step 42 except that they are derived from the texture image and the range image prerecorded depth. Typically the construction of maps J _ire f i for each direction is carried out before step 40 and are then stored in the database 20.

The intermediate score S _iT of similarity is a real number representative of the probability that the gradient maps J _iT and _Jire fT correspond to the same object. By example, here, the intermediate score S _iT is even larger than the probability that it is an image of the same object is important.

Here, each score S _iT is calculated according to the invariant scaling visual characteristic transformation method, better known by the acronym SIFT (Scale Invariant Feature Transform).

The SIFT method is well known and will therefore not be described in more detail. For more information on this method, the reader can refer to the patent application US 671 1293 and to the following article:

D. G. Lowe, "Distinctive Image Features from Scale-invariant Keypoints",

International Journal of Computer Vision, 60 (4): 91-1, 10, 2004.

J cards, highlight a greater number of key points or characteristic points of an object than the image initially acquired in step 40. But these are the key points that are compared to the key points of the map _fire f to establish the similarity score between these maps. For example, it has been measured that the SIFT method identifies 41 and 61 key points respectively in telemetry and texture images while the same method identifies 1 16 and 304 key points in the corresponding "JirefR" and "JirefT" maps.

Here, the steps 42 and 52 are also performed for the telemetric depth image acquired during step 40 so as to obtain intermediate scores S _iR of similarity between the cards J _iR and J _ire fR.

Once the intermediate scores S _iT and S _{iR have been} calculated, during a step 54, the calculator 16 establishes a final score S _f of similarity from these intermediate scores S _iT and S _iR . Here, the score S _f is established using a weighted sum of scores S _iT and S _iR . Thus, the score S _f is given by the following relation:

i = N

S _f = /. w S where w, is a predetermined weight.

[0071] Different methods can be used to determine the value of the weights w. For example, a genetic algorithm such as that described in the following document is advantageously implemented:

Y.H. Said, "On Genetic Algorithms and their Applications" Handbook of statistics,

2005.

Then, during a step 56, the computer 16 establishes the correspondence between the acquired face and the face contained in one of the prerecorded images. For example, if the final score S _f exceeds a predetermined threshold Bi when compared to one of the pre-recorded images, then the face of the user is considered recognized. In this case, a step 58 is performed in which the computer 16 authorizes access to the resource 4. For example, in the context of access to a building, the computer 16 controls the unlocking of a door. In the case where the final score S _f does not cross the threshold Bi, the user is considered not to have been recognized and we return to step 40.

The previously described method was tested using a standard database known by the acronym FRGC, version 2, so as to allow a comparison of the results obtained with other known methods. The FRGC Face Database, version 2, is well known. For example, it is presented in the following article:

PJ Phillips, PJ Flynn, Scruggs T., KW Bowyers, K. I Chang, K. Hoffman, J. Marques, J. Min and Worek W., "Overview of the face recognition great challenge" International Conference on Computer Vision and Pattern Recognition, 2005.

This database contains 4007 3D faces, each 3D face containing both intensity values and depth telemetry for each pixel. Thus, for each 3D face, one has a texture image and a telemetric depth image.

Pretreatment has been applied to the images of this database for:

to eliminate the abnormally high variations of the value a (x, y) in the image by using a median filter, and

- Fill in the holes using least squares interpolation from the values of the adjacent pixels.

Of the 4007 3D faces, 466 were retained to form the J _iref cards stored in the database 20. The remaining 3541 3D faces were used for the test.

The following table gives the recognition rate ("rank-one rate" in English) and the verification rate with a FAR (False Acceptance Resuit) of 0, 1% obtained by applying different processes, including the method previously described, to the FRGC database.

Method Used Rate Rate

recognition verification with

(rank-one rate) a FAR of 0, 1%

Mian et al [24] 0.961 0.986

Mian et al [23] 0.974 0.993

Gokberk et al [25] 0.955 NA

Xu et al [26] NA 0.975

Maurer et al [13] NA 0,958

Husken et al [1 1] NA 0.973

Ben Soltana et al [28] 0.955 0.970

The process described above 0.980 0.989 The symbol NA in the table means that the data are not available.

The references indicated in square brackets in the table above correspond to the methods described in the following articles:

- [24] A.S. Mian, M. Bennamoun and R. Owens, "Keypoint Detection and Local Feature Matching for Textured 3D Face Recognition," International Journal of Computer Vision, 79 (1): 1-12, 2008,

- [23] AS Mian, M. Bennamoun and R. Owens, "An Efficient Multimodal 2D-3D Hybrid Approach to Automatic Face Recognition", IEEE Transactions in Pattern Analysis and Machine Intelligence, 29 (1 1): 1927-1943, 2007 ,

- [25] B. Gokberk, H. Dutagaci, A. Ulas, L. Akarun and B. Sankur, "Representation Plurality and Fusion for 3D Face Recognition", IEEE Transactions on Systems, Man and Cybernetics - B: Cybernetics, 38 ( 1): 155-173, 2008,

- [26] C. Xu, F. Li, T. Tan and L. Quan, "Automatic 3D Face Recognition from Depth and Intensity Gabor Features", Pattern Recognition, 42 (9): 1895-1905, 2009,

- [13] T. Maurer et al, "Performance of Geometrix ActivelD 3D Face Recognition Engine on FRGC Data", IEEE Workshop on FRGC Experiments, 2005,

- [1 1] M. Husken, M. Brauckmann, S. Gehlen and C. v. Malsburg, "Strategies and Benefits of Fusion of 2D and 3D Face Recognition", IEEE Workshop on Face Recognition Grand Challenge Experiments, 2005,

- [28] W. Ben Soltana, D. Huang, M. Ardabilian, L. Chen and C. Ben Amar, "Comparison of 2D / 3D Features and their Adaptive Level Fusion Score for 3D Face Recognition", 3D Data Processing, Visualization and Transmission, 2010.

As shown in the table above, the method described here obtains the best recognition rate among the various processes to which it has been compared.

Many other embodiments are possible. For example, other sensors than those described can be used to acquire an image. For example, the sensor may be a digital camera or a scanner. The electronic sensor may also be a computer on which software for capturing or generating an image is executed in response to commands from a user. For example, the sensor may be a computer on which CAD (Computer Aided Drawing) software is running. Indeed, such a CAD software makes it possible to acquire an image in response to the commands of the user. In the latter case, the values a (x, y) of each pixel are not measured but generated by the software. The number of directions used to construct the different gradient maps may be different from eight. In addition, the different directions used can be chosen so that they are not uniformly distributed in the plane of the image. In this case, the angle Θ between two immediately successive directions is not necessarily the same between all the successive directions. Finally, the number of directions taken into account for the texture image is not necessarily the same as that used for the telemetric depth image. The directions used for the texture image are not necessarily the same as those used for the telemetric depth image.

In the cards, it is also possible to keep only the negative values and to replace the positive values with a zero value. If in the gradient maps the negative and positive values are kept, the number H of direction sets can be divided by two.

The method may also comprise the determination of the direction Oi to be used for each pixel P (x, y) so as to make these cards independent of the rotation of an image. Since in this embodiment, the direction Oi depends on the pixel concerned, it is denoted Oi (x, y). For this, the direction Oi (x, y) of each pixel is chosen as the main direction of the gradients of the neighboring pixels. The principal direction of the gradients of the neighboring pixels is calculated as the average of the gradients of the set of pixels in the neighborhood of a pixel. These pixels are considered as neighbors if they are contained within a predetermined perimeter centered on the pixel for which its direction Oi is calculated. This perimeter is for example a circle of radius R chosen. For example, initially, the gradients in the horizontal and vertical directions are calculated for each pixel of the image. We denote g (x, y) the gradient vector of each pixel P (x, y) whose components are formed by the horizontal and vertical gradients calculated for this pixel P (x, y). The principal direction of the gradients of the neighboring pixels is, for example, determined by summing the vectors g (x, y) of neighboring pixels and, possibly, dividing this sum by A, where A is the number of neighboring pixels. Thus, in this embodiment, the direction Oi (x, y) of one pixel may be different from the direction Oi (x ', y') of another pixel of the same image. Then, the method described above is applied for each pixel P (x, y) of the image but using the direction Oi (x, y) determined for this pixel and not a direction Oi common to all pixels of the image. The directions O ₂ (x, y), O ₃ (x, y) and following associated with the pixel P (x, y) are constructed from the direction Oi (x, y). For example, the direction Oi (x, y) is deduced from the direction Oi (x, y) by applying an angular rotation of a predetermined angle Θ. The angle Θ, which makes it possible to construct the direction Ο, (χ, ν) from the direction Oi (x, y) is the same regardless of the pixel P (x, y) of the image. Therefore, in this embodiment, the index "i" does not identify a direction Oi common to all the pixels of the image but a set i of directions Ο, (χ, ν) associated with each pixel P (x, y) of the image. Such a choice of the direction Oi (x, y) makes it possible to make the above method robust with respect to the rotation of the object around an axis perpendicular to the plane of the image. The different variants described in the particular case where all the directions Ο, (χ, ν) of the set i are parallel, also apply to the case where the directions Ο, (χ, ν) vary from one pixel to another of the same picture. In another embodiment, the database 20 contains only the cards J _ire f without the texture images and the corresponding telemetric images.

The method described above can also be performed only on the texture images or only on depth telemetry images. In the case where only the texture images are used, we will speak of 2D recognition.

It is also possible that the number 0 _T of gradient maps that are calculated on a texture image is different from the number _R 0 of gradient maps that are calculated on the depth telemetric image. In this case, N is the sum of 0 _T and

Other formulas for calculating the gradient of a pixel are possible. For example, the following formula, given for the direction Oi, can also be used: Pi (x, y) = a (x, y) - a (x + 1, y). Other examples of formulas are also given in patent application US 2010 0135540, paragraphs 46 et seq.

[0090] Other methods of smoothing the gradient map can be implemented. It is also possible not to perform smoothing.

Other methods of normalizing the vector of directional gradients are possible. For example, the vector p ^R (x, y) can be normalized by dividing its components by the smallest value of these or more generally a linear combination of these as for example the sum of its components. Thus, in a variant, this normalization consists in dividing each value Pi _T ^R (x, y) by the maximum or the minimum of the components of the vector p _T ^R (x, y). In case the vector pr ^R (x, y) is normalized by the maximum of its components, p _iT ^N (x, y) is given by the following relation: p _iT ^N (x, y) = p _iT ^R (x , y) / Max [pi _T ^R (x, y); ...; p _H T ^R (x, y)], where "Max" denotes the function that returns the maximum of the various components contained in square brackets.

In another variant, the normalization operation is omitted.

Other methods of extracting features than the SIFT method can be used. For example, the DAISY method can also be used. This method is for example described in the following article:

E. Tola, V. Lepetit and P. Fua, "A Fast Local Descriptor for Dense Matching," IEEE International Conference on Computer Vision and Pattern Recognition, 2008.

In the weighted sum used to establish the final score S _f of similarity, all the weights w can have the same value.

An object can be considered recognized when a correspondence has been established between the image of this object and the reference image with which it has the highest final similarity score. In this case, it is not necessary that the final score exceeds a predetermined threshold for recognition of the object. In a variant, the prerecorded image library may contain only one image.

The resource 4 is not necessarily a gateway to a building, it can be all types of resources such as a computer or software. The biometric system may also configure parameters of the resource 4 based on the identified person.

The process for establishing the final similarity score can be used in other contexts than a method for recognizing an object. For example, it can be used to classify or search by similarity images between them without necessarily trying to identify the object contained in these images.

The object may be something other than a face. For example, in the context of a method of recognizing a human being, it may be a fingerprint, the veins of a hand when presenting the fist closed in front of the electronic sensor, the iris with one eye. The object can also be something other than a biometric feature. For example, the object may be the texture of a material such as the texture of a paper, a textile or a bill. The object may also be a landscape or the like when the method for establishing the final similarity score is implemented, for example, in similarity image search software.

Claims

1. A method of establishing a final score of similarity between a digital image of an object, acquired by an electronic sensor, and a prerecorded digital image in an image library, each digital image being formed of pixels P (x, y ) and having a matrix which at each position (x, y) of a pixel in the image associates at least one value a (x, y) of a physical quantity acquired by the electronic sensor, which method comprises:

a) the construction (42) from the acquired image of H maps J, gradients calculated according to a respective set of directions, where:

H is an integer, of games i of directions, greater than or equal to two,

a set i of directions associates with each pixel P (x, y) of the image a direction Ο, (χ, ν), the direction Ο, (χ, ν) being angularly shifted with respect to the direction Oi (x , y) of an angle θ ,, this angle Θ, being the same regardless of the pixel considered in the image,

a map of gradients calculated according to a set of directions being a matrix which at each pixel P (x, y) of the acquired image associates a value ρ, (χ, ν) representative of the gradient of the physical quantity in the direction Ο, (χ, ν) associated with this pixel in the game i,

characterized in that the method comprises:

b) for each gradient map J ,, the computation (52) of an intermediate score S, of similarity between this gradient map J, and a corresponding reference map J _ire f, each intermediate score S, of similarity being a a real number representing the degree of similarity between the compared maps, the Jiref reference map being a map of gradients calculated according to the same set of directions as the gradient map J, from the prerecorded image, and

c) establishing (54) a final score S _f of similarity between the acquired image and the prerecorded image as a function of the H intermediate scores S, of similarity.

The method of claim 1, wherein the J-map construction of gradients comprises normalizing (48) the gradient associated with each pixel of said map by dividing it by at least one component or a combination of components of a vector of directional gradients, this component or combination of components being such that if each component of the vector of directional gradients and this gradient are multiplied by the same constant λ, the normalized gradient remains independent of this constant λ, the vector of directional gradients gathering only gradient values, in different directions, associated with this same pixel.

The method of claim 2, wherein the J-map construction of gradients comprises normalizing (48) the gradient associated with each pixel of said map by dividing it by the directional gradient vector standard or the maximum or minimum or a linear combination of the components of this same vector of directional gradients.

A method according to any one of the preceding claims, wherein the method comprises determining the direction Oi (x, y) associated with each pixel P (x, y) in the set of directions, from the gradients, in at least two different directions, neighboring pixels included in a predetermined perimeter centered on this pixel P (x, y).

5. Method according to any one of claims 1 to 3, wherein the direction Oi (x, y) associated with each pixel P (x, y) in the set of directions is the same regardless of the pixel of the 'picture.

6. Method according to any one of the preceding claims, wherein the construction of each map J of gradients comprises smoothing (46) the gradient associated with each pixel, said central pixel, of this map by realizing a weighted average of the gradient. of the central pixel with those of the adjacent pixels located within a predetermined perimeter centered on the central pixel.

7. The method according to claim 6, wherein weighting coefficients are assigned to each gradient of one pixel of the weighted average, the weighting coefficient assigned to a gradient of an adjacent pixel being all the smaller as the pixel. adjacent is far from the central pixel.

A method according to any one of the preceding claims, wherein steps a) and b) are performed on a texture image and on a depth telemetric image of the object and, in step c), the final score S _f is established from the intermediate scores S, obtained from the texture image and the telemetric depth image.

9. A method according to any one of the preceding claims, wherein the establishment of the final score S _f comprises performing a weighted sum of intermediate scores S ,, a different weight being assigned to each intermediate score of this sum.

The method of claim 9, wherein the weights of the weighted sum are selected by implementing a genetic algorithm.

1 1. A method as claimed in any one of the preceding claims, wherein H is an integer greater than or equal to two if the gradient maps have negative and positive values or four if the gradient maps comprise only null or positive values.

12. Method for recognizing an object, this method comprising:

the acquisition (40) by an electronic sensor of a digital image of the object to be recognized, each digital image being formed of pixels P (x, y) and comprising a matrix which associates with each position (x, y) a pixel P (x, y) in the image at least one value a (x, y) of a physical quantity measured by the electronic sensor,

characterized in that the method comprises:

calculating a final score S _f of similarity between the acquired image and each prerecorded image of a digital image library by implementing a method according to any one of the preceding claims, and

the establishment (56) of the correspondence between the acquired image and one of the prerecorded images of the library as a function of the final score S _f of similarity calculated between these two images, the object being recognized when the acquired image of this object corresponds to one of the pre-recorded images.

13. Information recording medium, characterized in that it comprises instructions for the implementation of a method according to any one of the preceding claims, when these instructions are executed by an electronic computer.

14. Device for recognizing an object, this device comprising:

an electronic sensor (10, 12) capable of acquiring a digital image of the object to be recognized, each digital image being formed of pixels P (x, y) and comprising a matrix that associates with each position (x, y) d a pixel in the image at least one value a (x, y) of a physical quantity measured by the electronic sensor,

an electronic calculator (16) programmed to construct, from the acquired image, H maps J, gradients calculated according to a respective set of directions, where:

H is an integer, of games i of directions, greater than or equal to two,

i is an identifier of the set of directions in which the gradients are calculated, a set i of directions associates with each pixel P (x, y) of the image a direction Oi (x, y), the direction Ο, (χ, ν) being angularly shifted with respect to the direction Oi (x, y) an angle θ ,, this angle Θ, being the same regardless of the pixel considered in the image,

a map of gradients calculated according to a set of directions being a matrix which at each pixel P (x, y) of the acquired image associates a value ρ, (χ, γ) representative of the gradient of the physical quantity in the direction Oi (x, y) associated with this pixel in the game i,

characterized in that the electronic computer (16) is also programmed to:

a) for each gradient map J ,, calculate an intermediate score S, of similarity between this gradient map J, and a corresponding reference map J _ire f, each intermediate score S, of similarity being a real number representing the degree of similarity between the compared maps, the reference map J _ire f being a map of gradients calculated according to the same set of directions as the gradient map J, from a prerecorded image of a digital image library, and

b) establishing a final score S _f of similarity between the acquired image and the prerecorded image as a function of the H intermediate scores S, of similarity,

c) repeating steps a) and b) for each prerecorded image of the library and establishing the correspondence between the acquired image and one of the prerecorded images of the library according to the final score S _f of similarity calculated between these two images , the object being recognized when the acquired image of this object corresponds to one of the prerecorded images.