Method of classification of an input image representative of a biometric trait by means of a convolutional neural network

1. A method of classification of an input image representative of a biometric trait by means of a first convolutional neural network, CNN, characterized in that it comprises the implementation, by data processing means ( 21 ) of a client ( 2 ), of steps of:

(a) estimation of a vector of parameters that are descriptive of a singular point of the biometric trait in said input image, by means of a third CNN;

(b) reframing said input image based on estimated parameters of said singular point, so that the parameters of said singular point have predetermined values for the reframed input image;

(c) estimation of a transformation parameter vector of said input image, by means of a second CNN, the parameters of the vector being representative of a geometric transformation enabling the biometric trait represented by the input image to be registered in a common reference frame;

(d) application to said reframed input image of a transformation defined by said estimated transformation parameter vector, so as to obtain a registered input image;

(e) classification of the registered input image by means of the first CNN.

2. The method according to claim 1 , wherein said biometric traits are chosen from among fingerprints, faces and irises, particularly fingerprints.

3. The method according to claim 1 , wherein said transformation parameters comprise at least one set of deformation coefficients each associated with a reference nonlinear deformation function from a family of reference nonlinear deformation functions, said transformation defined by said estimated transformation parameters vector comprising a deformation expressed from said family of reference nonlinear deformation functions and associated coefficients.

4. The method according to claim 3 , wherein said reference nonlinear deformation functions are velocity fields (E 1 ), said deformation being expressed in the form of a diffeomorphic distortion field (T) as an exponential of a linear combination (T=exp(Σ i=0 k-1 c i E i )) of the velocity fields (E i ) weighted by said associated coefficients (c i ).

5. The method according to claim 3 , wherein said transformation parameters further comprise a rotation parameter and/or a change of scale parameter and/or at least one translation parameter.

6. The method according to claim 5 , wherein said transformation defined by said estimated transformation parameters vector comprises a composition of said deformation with an affine transformation expressed from corresponding transformation parameters.

7. The method according to claim 1 , wherein said transformation parameters comprise at least one coordinate of the singular point and an angle of the singular point, said reframing of the input image comprising a translation and/or a rotation.

8. The method according to claim 1 , comprising a prior training step (a0), by data processing means ( 11 ) of a server ( 1 ), from a database of training images already classified, from parameters of said first and second CNNs.

9. The method according to claim 8 , wherein the first and second CNNs are trained simultaneously and semi-supervised, the training images from the database of training images not being associated with transformation parameters.

10. The method according to claim 1 , wherein the step (a0) also comprises the training of the third CNN, the training images from the database of training images being associated with parameters that are descriptive of a singular point.

11. The method according to claim 9 , wherein the training of the first and second CNNs comprises, for at least one training image from said training image database, obtaining a registered training image, classifying the registered training image, and minimizing a loss function.

12. The method according to claim 8 , wherein the step (a0) preliminarily comprises the statistical analysis of said database of already classified training images, so as to determine said family of reference nonlinear deformation functions by which transformations observed in the training database can be expressed based on a set of coefficients.

13. The method according to claim 12 , wherein said transformations observed in the training database are nonlinear transformations with which to map from one member of a training data pair to the other where the pair is representative of a single biometric trait.

14. The method according to claim 12 , wherein said statistical analysis is a main component analysis on all the distortion fields expressing the transformations observed in the training database, the reference nonlinear deformation functions being determined as velocity fields defined by eigenvectors produced from the main component analysis.

15. The method according to claim 1 , wherein said biometric trait represented by the input image is that of an individual, step (e) being a step of identification or authentication of said individual.

16. A non-transitory computer readable program product comprising code instructions for the execution of a method according to claim 1 of classification of an input image representative of a biometric trait by means of a first convolutional neural network, CNN, when said program is executed by a computer.

17. A non-transitory storage means readable by computer equipment on which a computer program product comprises code instructions for the execution of a method according to claim 1 of classification of an input image representative of a biometric trait by means of a convolutional neural network, CNN.