Parameter training method for a convolutional neural network

A parameter training method for a convolutional neural network (CNN) for classifying image type data representative of a biometric trait, comprising an implementation, by data processing means of a server, the steps of:

• a) For at least one data item from a training database which stores already classified training data, generating several alternate versions of the at least one data item each by application of at least one geometric transformation chosen from a set of reference transformations satisfying a statistical distribution of geometric transformations observed in the training database, comprising a construction of a statistical atlas of the training database which includes an average model of images from the training database and a variability of geometric transformations observed in the training database;
• b) Training the parameters of said CNN, from the training database augmented with said alternate versions of the at least one data item:
• wherein step a) comprises substeps of:
• a1) Statistical modeling of the transformations observed in said training database by means of said statistical atlas;
• a2) Determination of said set of reference transformations satisfying a statistical distribution of transformations observed in the training database depending on the result of said statistical modeling;
• 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, wherein step a1) comprises the determination of the observed transformations;
• and
• wherein said average model of images comprises, for each set of already classified training data representative of a single biometric trait, an average image associated with said set the step a1) further comprising the determination of nonlinear relative transformations with which to map an already classified training data to the average image associated with the set to which said already classified training data belongs;
• the statistical modeling of the transformations observed in said training database being done on the basis of said nonlinear relative transformations.