Method and system for classifying fingerprints

Abstract

Method (3000), implemented by computer, for classifying handwriting into a plurality of membership classes C_j, each of which corresponds to a particular anatomical area Z_i of the palmar surface of a hand, said method (2000) takes, as input data (I3000), at least one handwriting D, and provides, as output data (O3000), a membership class C or a list L of membership classes C_k of the handwriting D from among the plurality of membership classes C_j.

Description

Title of the invention: Method and system for classifying typescripts. Technical field.

[0001] The invention relates to a method and system for classifying fingerprints. The invention allows for automatic identification of the type of fingerprints acquired by a fingerprint acquisition device. Technical background

[0002] Fingerprints, generally known as "fingerprints" and / or "palm prints," are patterns formed by the traces left by dermatoglyphs of the fingers and / or palms on surfaces. Dermatoglyphs are the superficial furrows formed on the palms, soles, and fingertips by the dermal ridges and arranged in lines or spirals. They are unique to each individual, and the patterns they form constitute a biometric "identity card" by which the individual can be identified.

[0003] Fingerprint scanning is a common practice in various administrative procedures with government institutions and in operations conducted by law enforcement agencies with a suspect or defendant in connection with an offense, misdemeanor, or felony. This scanning is generally carried out using appropriate fingerprint acquisition devices.

[0004] These acquisition devices generally consist of an electronic box comprising an acquisition surface equipped with a sensor on which one or more fingers, the hand or a part of the hand are placed to acquire an image of the dermatoglyphs.

[0005] US 2012 014569 Al, IB KOREA LTD [KR], 19.01.2012 describes a portable device for acquiring handwriting samples. The device comprises an acquisition surface equipped with an electroluminescent sensor onto which the fingers of a hand can be placed to acquire an image of their dermatoglyphs.

[0006] US 2017 046554 A1, NEC CORP [JP], 16.02.2017 describes a portable device for acquiring digital fingerprints. This device is configured to provide instructions enabling a user to correctly acquire fingerprints.

[0007] Today, fingerprints are biometric information widely used to identify individuals and / or authenticate a transaction such as an online banking transaction or an individual's use of a password from those stored in a password wallet. However, unlike authentication where a fingerprint acquired for an individual is Compared to a single or very limited number of reference fingerprints (1:1), identifying an individual from a fingerprint requires comparing that fingerprint with many fingerprints previously acquired from several individuals (1:N) and usually stored in a database.

[0008] Because handwriting samples are drawings with complex characteristics and the number of comparisons required for identification can become very high, comparing dermatoglyphs can be time-consuming despite the computing power of currently available data processing devices. To reduce the time required for these operations, it is known to classify handwriting samples into different classes based on certain morphological characteristics of the dermatoglyphs. For example, these morphological characteristics may include the general shape of the dermatoglyph (orientation of loops, arches, spirals, etc.), notably according to the categories of Henry Faulds, Francis Galton, and Edward Henry; the overall shape of the ridges; the "minutiae" consisting of singular points along the ridges (termination of a ridge, bifurcation, etc.); the shape of the ridges; pores; or even scars.

[0009] US 5572597 A1, LORAL CORP [US], 05.11.1996 describes a method for classifying fingerprints into different types based on the overall patterns formed by dermatoglyphs, including ridges and furrows, across the entire fingerprint. The classification method is based on an analysis of the directions and angles formed by the local patterns of ridges and furrows in small areas of interest within the fingerprints.

[0010] EP 0 779 595 A2, NEC CORP [JP], 18.06.1997 describes a method for classifying fingerprints into five categories derived from the categories of Henry Faulds, Francis Galton, and Edward Henry: simple arc, extended arc, right loop, left loop, and spiral. The method is based on a combination of the results of two classifiers to calculate a probability of a fingerprint belonging to one of these four categories.

[0011] US 5 825 907 A1, LUCENT TECHNOLOGIES INC [US], October 20, 1998, describes a method for classifying digital handwriting into five categories derived from the categories of Henry Faulds, Francis Galton, and Edward Henry: right loop, left loop, arc, spiral, and double loop. The method implements an artificial neural network configured to classify digital handwriting based on a map of the local shapes of the grooves.

[0012] Wang et al., Fingerprint Classification Based on Depth Neural Network, arXiv preprint arXiv: 1409.5188, 2014, describes a method for classifying fingerprints according to Edward Henry's four categories: arc, right loop, left loop, and spiral. The method implements, in a In the first step, an artificial neural network trained using unsupervised learning extracts the orientation map of a digital fingerprint—the orientation map of a digital fingerprint corresponds to the general pattern of direction of the dermatoglyph's ridges. In the second step, a logical regression model trained using supervised learning for fuzzy classification calculates the probability that the orientation map belongs to one of four categories.

[0013] US 9 530 042 Bl, UNIV KING SAUD [SA], 27.12.2016 describes a method for classifying digital fingerprints into four categories derived from the categories of Henry Faulds, Francis Galton, and Edward Henry: right loop, left loop, arc, and spiral. The method first performs a calculation of a feature vector for each digital fingerprint using a descriptor based on a local gradient directional binary model. The vectors are then processed by an artificial neural network configured to classify the fingerprints into one of the four categories.

[0014] Michelsanti et al., Fast Fingerprint Classification with Deep Neural Networks, VISIGRAPP 2017, describes a method for classifying fingerprints according to Edward Henry's four categories: arc, right loop, left loop, and spiral. The method implements a convolutional neural network of artificial neurons of the VGG-F or VGG-S type trained on a set of fingerprints classified according to the four categories. This method allows for the direct extraction of fingerprint features and thus eliminates the need for an intermediate extraction step.

[0015] EP 3 825 915 Al, IDEMIA IDENTITY & SECURITY FRANCE [FR], 26.05.2021 describes a method for classifying digital fingerprints according to a given number of categories derived from the categories of Henry Faulds, Francis Galton, and Edward Henry. The method implements a convolutional neural network of artificial neurons trained to determine whether or not a digital fingerprint belongs to each category. A fingerprint may belong to more than one category. Summary of the invention Technical problem

[0016] Depending on the context of operation and use—for example, a civil context for administrative procedures and / or border crossings; a criminal context for recording the fingerprints of a suspect or defendant—the acquisition of types of fingerprints, in addition to digital fingerprints, may be relevant because they contain additional biometric information that may be used as more precise and supplementary means of identifying or authenticating an individual. Also, many devices, such as those described above, are arranged to acquire, in addition to digital fingerprints, other types of fingerprints such as complete and / or partial palm fingerprints, or even fingerprints of several phalanges or several fingers.

[0017] As with digital fingerprints, their classification is a prerequisite for the efficient subsequent use of these other types of fingerprints. Furthermore, because they are generally acquired on the same device during the same acquisition campaign, this classification must at least distinguish between these different types. However, the classification methods for digital fingerprints are completely unsuitable for such an operation because they exclude all fingerprints other than digital ones.

[0018] A first negative consequence is that it falls to the operator to distinguish or classify the different types of typescripts during the acquisition campaign. However conscientious the operator may be, the risk of error remains, both in correctly assigning the typescript to a category and in identifying typescripts that do not conform to the expected type during acquisition.

[0019] A second negative consequence is the equally obvious inadequacy of the aforementioned current methods for verifying the integrity, namely the accuracy, completeness, and reliability, of the fingerprints in a database comprising several types of fingerprints, and, where appropriate, for proposing corrections. More specifically, these methods can only verify that a partial palm fingerprint corresponding to the lower part of a palm has not been mistakenly or inadvertently classified as a fingerprint of several fingers or phalanges.

[0020] There is therefore a need for a reliable and automatic method of identification and / or classification of different types of handwriting acquired during multimodal acquisition campaigns and / or using, in particular, multimodal devices such as those described above. Technical solution

[0021] In a first aspect of the invention, a computer-implemented method is provided for classifying fingerprints into a plurality of membership classes, each class corresponding to a particular anatomical area of ​​the palmar surface of a hand. The method takes at least one fingerprint as input data and provides, as output data, a membership class or a list of membership classes for the fingerprint from among the plurality of membership classes. The method comprises the following steps: - provide a convolutional network of artificial neurons, said convolutional network being previously trained on a training dataset composed of a plurality of fingerprints classified according to a plurality of classes, each of which corresponds to a particular anatomical area of ​​a hand, said convolutional network being configured to provide the probabilities of each fingerprint in set E belonging to each of the classes of the plurality of membership classes; - infer, using the trained convolutional neural network, the probabilities of belonging of the typescript provided as input data to each of the classes of the plurality of membership classes; - select a membership class from among the membership classes for which the membership probability of the typescript provided as input data is the highest among the membership probabilities inferred for said classes, or a list of a number of membership classes selected from the plurality of membership classes and sorted in ascending or descending order of the membership probabilities in said classes inferred for the typescript provided as input data.

[0022] Advantageous embodiments of the first aspect of the invention are described in detail below.

[0023] In a second aspect of the invention, a data processing device is provided comprising means for implementing a method according to the first aspect of the invention.

[0024] In a third aspect of the invention, a computer program is provided comprising instructions which, when the program is executed by a computer, lead the computer to implement a method according to the first aspect of the invention.

[0025] In a fourth aspect of the invention, a computer-readable medium is provided comprising instructions which, when executed by a computer, lead the computer to implement a method according to the first aspect of the invention.

[0026] In a fifth aspect of the invention, a classification system is provided among a plurality of membership classes, said system comprising a device for acquiring typescripts and a data processing device according to the second aspect of the invention. Brief description of the drawings

[0027] [Fig.1] is an example of a device for acquiring typescripts.

[0028] [Fig.2] is a table grouping examples of schematic representations of classes of typing diagrams in which each of the classes corresponds to a particular anatomical area of ​​a hand.

[0029] [Fig.3] is a flowchart of a method according to the first aspect of the invention.

[0030] [Fig.4] is a schematic representation of the architecture of a convolutional network of artificial neurons according to an example of an embodiment of a method according to the first aspect of the invention.

[0031] [Fig.5] is a representation of a data processing device according to the second aspect of the invention.

[0032] [Fig.6] is a representation of the classification performance of an example of a method according to the first aspect of the invention expressed as the number of classification successes and failures as a function of the probability of belonging to the class predicted by said method. Detailed description of the implementation methods

[0033] Figure 1 schematically illustrates an example of a multimodal fingerprint acquisition device 1000, capable of acquiring not only digital fingerprints but also other types of fingerprints, such as complete and / or partial palm fingerprints, or fingerprints of several phalanges or fingers. The device generally comprises an electronic unit 1001 equipped with an acquisition surface 1002 onto which one or more fingers, the hand, or a part of the hand can be placed to acquire an image of the dermatoglyphs. Numerous devices of this type are described in the prior art.

[0034] Fingerprints acquired by a multimodal device such as the one described above are generally presented in the form of an image of the designs formed by the dermatoglyphs. These images can be single-channel images, for example, a greyscale image, or multi-channel images, for example, RGB images.

[0035] Fingerprints can be classified into several classes, each class corresponding to a particular anatomical area of ​​the palmar surface of the hand to which they relate. The "palmar surface" of the hand is understood to be the surface of the hand that includes the palm, as opposed to the dorsal surface. The palmar surface includes the palm and all the fingers, that is, the thumb, index, middle, ring, and little fingers, the thenar eminence, the hypothenar eminence, and the hypothenar eminence. This definition corresponds to that commonly accepted in anatomy.

[0036] The number of membership classes is not limited and can vary according to the needs and contexts of operation and use. Similarly, the nature of the classes their area of ​​belonging, that is to say the anatomical zones of the palmar surface of the hand to which they correspond, can be defined or customized according to the same needs and contexts of operation and use

[0037] Nevertheless, in practice, with reference to [Fig.2], it can be distinguished, among the different types of fingerprints, 11 classes Cl-Cll of belonging corresponding to different anatomical areas of the palmar surface of the hand and adapted to a very large number of needs and use cases: the complete right hand (Cl), the complete left hand (C2), a lower area of ​​the palm of the right hand (C3), a lower area of ​​the palm of the left hand (C4), a higher area of ​​the right hand (C5), a higher area of ​​the left hand (C6), the writer's right palm (C7), the writer's left palm (C8), at least two, preferably three, preferably four fingers of the right hand (C9), at least two, preferably three, preferably four fingers of the left hand (C10), the thumbs of the right and left hands (C11).

[0038] By way of example, in a first registration framework for administrative procedures, it may be advantageous to acquire five different classes such as the lower area of ​​the right palm, the lower area of ​​the left palm, four fingers of the right hand, four fingers of the left hand, and both thumbs. Conversely, in a second registration framework for the fingerprints of a suspect or defendant, the number of fingerprint types, and therefore the number of classes, may be higher in order to obtain the most comprehensive biometric information possible. In addition to the previous classes, it includes one or more classes from among the entire right hand, the entire left hand, the upper area of ​​the right hand, the upper area of ​​the left hand, the writer's right palm, and the writer's left palm.

[0039] As explained previously, for efficient use and / or effective acquisition of the different types of fingerprints that can be acquired, it is necessary to classify them. However, current methods for classifying digital fingerprints are completely inadequate, with the negative consequences described above.

[0040] Also, according to a first aspect of the invention, with reference to [Fig. 3], a computer-implemented method 3000 is provided for classifying fingerprints into a plurality of membership classes Cj, each of which corresponds to a particular anatomical area Z_i of the palmar surface of a hand. Method 3000 takes, as input data 13000, at least one fingerprint D, and provides, as output data 03000, a membership class C or a list L of membership classes C_k of the fingerprint D from among the plurality of membership classes C_j. Method 3000 comprises the following steps: - provide 3001 a convolutional neural network (CNN) of artificial neurons, said convolutional neural network being previously trained on a training dataset E consisting of a plurality of fingerprints D_i classified according to a plurality of classes Cj, each of which corresponds to a particular anatomical area of ​​a hand, said convolutional neural network being configured to provide the probabilities of membership Pj of each fingerprint D_i of the set E to each of the classes of the plurality of membership classes Cj; - infer 3002, using the trained convolutional neural network CNN, the membership probabilities P_i of the typescript D provided as input data 13000 to each of the classes of the plurality of membership classes Cj; - select 3003 a membership class C from among the membership classes Cj for which the membership probability P of the typescript D supplied as input data 13000 is the highest among the membership probabilities P_i inferred for said classes Cj, or a list L of a number k of membership classes C_k selected from the plurality of membership classes Cj and sorted in ascending or descending order of the membership probabilities P_k of said classes C_k inferred for the typescript D supplied as input data 13000.

[0041] The term “plurality of classes” means a set comprising at least two classes.

[0042] In practice, at inference step 3002, the output of the convolutional neural network CNN provides a vector or list V = {Pj}, each value of which is the probability Pj of membership for each of the classes in the plurality of classes Cj of membership. Thus, there is a vector V for each typescript D provided as input data 13000.

[0043] In step 3001, according to a first alternative, a membership class C is selected from among the membership classes Cj for which the membership probability P of the typescript D provided as input data 13000 is the highest among the membership probabilities P_i inferred for said classes C_j. This selection operation can be expressed according to the following mathematical formula: C=Q€{C J}, P^max^}

[0044] According to a second alternative, a list L of a number k of membership classes C_k is selected from the plurality of membership classes Cj and sorted according to a decreasing order of the probabilities P_k of membership in said classes C_k inferred for the typescript D provided as input data 13000.

[0045] A first notable advantage of the method according to the first aspect of the invention is that it reduces the risk of error that an operator might make when assigning a class to a typescript during or immediately after its acquisition. The method can thus be considered as providing a corrective, control, or double-checking function during acquisition.

[0046] According to a first implementation example, the method described in the first aspect can provide a function for checking the expected typeface class during a typeface acquisition campaign. By comparing the class it assigns to a newly acquired typeface to the expected class for that typeface, it allows for real-time verification that the acquisition campaign is proceeding according to a pre-established acquisition program, such as the acquisition of a single given typeface class or an order for acquiring typefaces of different classes. In case of non-compliance with the expected outcome, a warning signal can be sent to the operator by any appropriate means.For example, during an acquisition campaign, it may be expected that several fingerprints of different classes will be acquired in a specific order: four fingers of the right hand, then four fingers of the left hand, then the palm of the right hand, and finally the palm of the left hand. If, due to an error in the operator's instructions and / or an oversight by the person for whom the fingerprints are being acquired, a fingerprint of the left palm is acquired instead of the right, the method will, by comparing the class assigned to the acquired fingerprint to the expected class, send a warning signal to the operator so that they can perform a new acquisition.

[0047] According to a second implementation example, the method can provide a post-processing correction function. After the acquisition of a fingerprint for which an operator must manually assign a class, the method can verify that the class thus assigned by the operator is correct. For example, the method can verify that a fingerprint of a right palm has not been classified by the operator as a fingerprint of a left palm. In the event that the class initially assigned by the operator does not correspond to the one that the method according to the invention would have assigned, a warning signal can be issued by any appropriate means to the operator so that they can validate their classification or correct it based on the class suggested by the method.As before, this double verification, here first human and then machine, ensures the accuracy of the typing classification and the integrity of the databases in which the typings may subsequently be stored.

[0048] According to a third implementation example, the method can provide an a priori correction function. As soon as a typescript is acquired using a device, it can propose a class to which it belongs, or even several classes sorted in descending order of their probability of belonging, to an operator who has the choice of validating or rejecting this proposal, or of choosing a class from among those proposed. The risk of misclassification by the operator is thus reduced since they only have to choose one class or a limited number of them. Furthermore, particularly in the event that the typescript is acquired under difficult conditions, the risk of misclassification by the method itself can be eliminated by the intervention of the operator, who is responsible for the final decision regarding the choice of the assignment class.This double verification, first machine-based and then human-based, ensures the accuracy of the typing classification and the integrity of the databases in which the typings may subsequently be stored.

[0049] A second notable advantage of the method according to the first aspect of the invention is that it allows the integrity of a database of typescripts of different classes to be verified. For example, a database containing typescripts of different classes may contain classification errors. The method can advantageously be used to detect and correct these errors by providing it with each typescript from the database as input data and comparing the class it assigns to it based on the highest probability of belonging with the class entered in the database for that typescript. In the event of a mismatch between the two classes, automatic correction can be performed and / or the typescript can be referenced in an appropriate list for subsequent verification by an operator.

[0050] According to some preferred embodiments, with reference to [Fig. 3], the classes of the plurality of membership classes are selected from the entire right hand (C1), the entire left hand (C2), a lower area of ​​the palm of the right hand (C3), a lower area of ​​the palm of the left hand (C4), an upper area of ​​the right hand (C5), an upper area of ​​the left hand (C6), the writer's right palm (C7), the writer's left palm (C8), at least two, preferably three, preferably four fingers of the right hand (C9), at least two, preferably three, preferably four fingers of the left hand (C10), and the thumbs of the right and left hands (C11). The plurality of membership classes can thus comprise a number of classes equal to or greater than two, up to a total of eleven membership classes.

[0051] The convolutional network of artificial neurons is of any type suitable for implementing the method according to the first aspect of the invention. According to some Preferred implementations include the convolutional neural network (CNN) of artificial neurons: - a first SI sequence of convolutional layers of artificial neurons; - a second S2 sequence of residual convolutional blocks of artificial neurons; - a third S3 sequence comprising at least one layer of fully connected artificial neurons.

[0052] The first SI sequence of convolutional layers serves, on the one hand, to reduce the dimensions of the data type and, on the other hand, to increase the number of topographic maps (or feature maps) relevant for extracting low-level patterns and features such as color distribution and intensity, textures, shapes, and local contrasts. The second S2 sequence of convolutional blocks serves to increase the receptive field of the network while improving the extraction of low- and high-level features through the combination of topographic maps from different layers via residual connections.The final S3 sequence of fully connected artificial neuron layers functions to combine all the features extracted by the two previous sequences to establish a probability of a typogram belonging to each of the classes in the plurality of membership classes.

[0053] As explained previously, at the output of the S3 sequence of artificial neural layers, each value of the vector V = {P_j} represents a probability Pj of membership for each of the classes in the plurality of classes Cj to which it belongs. These values ​​are specific to the convolutional network and are generally in the form of real numbers whose distribution differs from the usual probability distributions on an interval [-1;1] or [0;1]. Therefore, the vector V can be converted using a conversion function, perhaps, for example, of the sigmoid type or, preferably, of the "softmax" type.

[0054] Figure 4 illustrates a detailed example of a convolutional CNN adapted for classifying typos provided in the form of multichannel images such as RGB format with a resolution less than 1000x1000 pixels between 100 and 500 dpi into the 11 membership classes according to the embodiments described above. In detail, the convolutional CNN network comprises: - a first SI sequence of four successive convolutional layers whose kernel dimension k, number n of kernel instances, stride step s and padding p are respectively: (k = 5, n = 8, s = 2, p = 2), (k = 3, n = 16, s = 2, p = 1), (k = 3, n = 32, s = 2, p = 1) and (k = 3, n = 96, s = 2, p = 1), and each convolutional layer being followed by a ReLu type correction layer; - a second sequence S2 of three residual convolutional blocks of the type "Depthwise Separable Convolutions"; each of the blocks comprising two sub-blocks SB1, SB2 with a residual connection at the output of the second sub-block, followed by a ReLu type correction layer before the input of the next block; the first sub-block SB1 is composed of a convolutional layer with n = 96 kernel instances of dimension k = 3, a step size s = 1, and a filling size p=l, a batch normalization layer and a ReLu type correction layer; the second sub-block SB2 is composed of a convolutional layer with n = 96 kernel instances of dimension k = 3, a step size s = 1, and a filling size p=l, and a batch normalization layer; - a third sequence S3 comprising a layer of fully connected artificial neurons providing as output an 11-dimensional vector, each dimension of which corresponds to each of the 11 membership classes described in the [Fig.2]; optionally, at the output of the second sequence S2 and before entering the third sequence, the topographic maps may be subjected to a flattening operation.

[0055] The convolutional neural network (CNN) of artificial neurons shown in [Fig. 4] is an example and is not limiting in scope with respect to the present invention. The number of layers in each of the sequences described above, the dimensions of the convolution kernels, the number of their instances, and the step and padding values ​​can be adapted to the different needs and use cases of typescripts, particularly with regard to format, resolution, and image acquisition conditions.

[0056] According to certain complementary embodiments, the convolutional neural network (CNN) of artificial neurons further comprises, between the first sequence S1 and the second sequence S2, a spatial resampling step, preferably by bilinear interpolation. The function of this resampling step is to resize topological maps (or feature maps) obtained as output from the second sequence into fixed-size topological maps before their input to the third sequence, regardless of the size of the data types provided as input to the method or the convolutional neural network of artificial neurons. The fixed size to which the topographic maps are reduced corresponds to the number of input neurons in the third sequence comprising at least one layer of fully connected artificial neurons.

[0057] Thanks to this spatial resampling step, the method can take as input data typescripts of any size. It can therefore be advantageously implemented with any type of typescript acquisition device without the need to resize the typescripts they are likely to provide. It can also be used for classifying typescripts contained in a database whose file sizes are not homogeneous.

[0058] The training, also called learning, of the convolutional neural network can be implemented in any suitable manner. According to some advantageous embodiments, the cost function used during the training of the convolutional neural network (CNN) includes a cross-entropy cost function. Specifically, a cross-entropy function compares the probability distribution, q—in this case, the probabilities PJ of belonging to each of the membership classes Cj—predicted by a model (the convolutional neural network (CNN)), with a reference probability distribution, p, of a training dataset to which the model is applied—in this case, the probabilities, equal to 0 or 1, of belonging to the classes Cj of the handwriting samples in a training set comprising previously sorted handwriting samples.The model training is implemented iteratively by minimizing the divergence between the two distributions. An example of a cross-entropy type cost function, f, can be expressed as follows, with j the number of membership classes Cj: . f= H(p, q) = -^jPU^gqtj)

[0059] When, in accordance with certain embodiments described above, the convolutional network of artificial neurons comprises a sequence including at least one layer of fully connected artificial neurons, the cost function may further include a conversion function, from the real vectors at the output of said layer to a probability distribution over the interval [0; 1]. The conversion function may, for example, be of the sigmoid type or, preferably, of the "softmax" type. This function is applied before the cross-entropy function.

[0060] Depending on the needs and use cases, the method according to the first aspect of the invention makes it possible to differentiate between right-handed fingerprints and left-handed fingerprints, and vice versa. It is preferably sensitive to chirality, that is, to the axial symmetry relationship between one hand and the other, for example, between a right-handed fingerprint and a left-handed fingerprint. Thus, according to certain advantageous embodiments, the cost function used during the training of the convolutional neural network (CNN) includes a function sensitive to axial symmetry.

[0061] Such an axial symmetry sensitivity function can be implemented during training as follows. For each typescript in a training set, a typescript is first created. The symmetrical form is obtained by performing an axial symmetry operation on the said fingerprint. Each fingerprint and its symmetrical counterpart are then provided as training input to the convolutional neural network to calculate the probabilities of belonging to each of the classes in the plurality of membership classes. These membership probabilities are presented, at the network's output, as a vector whose values ​​correspond to the probability of belonging to each of the classes in the plurality of membership classes. Thus, there is a vector V[D_i] for each fingerprint D_i in the training set and a vector Vs[sD_i] for each corresponding symmetrical fingerprint sD_i.

[0062] Following this classification step, the vectors V[D_i] undergo a symmetry operation consisting of assigning the probability values ​​of a class corresponding to an anatomical area of ​​one hand, for example the right hand, to the corresponding class of the other hand, for example the left hand; the vector thus obtained is denoted S(V[D_i]). By way of illustrative example, for a plurality of classes including the classes "fingerprint of the palm of the right hand" and "fingerprint of the palm of the left hand", the probability of belonging to the class "fingerprint of the palm of the right hand" is assigned to the class "fingerprint of the palm of the left hand" and vice versa.

[0063] The sensitivity function can then be expressed as a Euclidean distance between the vector Vs[sD_i] and the vector S(V[D_i]), possibly weighted by a factor. Integrated into the previous cost function, the sensitivity function can be written as: f= -^jPU^ogqtj^+aL^VslsDj], S( V[D7]))

[0064] During the training of the convolutional neural network, this distance is minimized using the cost function. For this purpose, an Adam-type optimization algorithm can be used.

[0065] According to certain preferred embodiments, the convolutional neural network (CNN) of artificial neurons is trained using a plurality of sets of training handwriting samples, each set comprising handwriting samples of identical dimensions. Thus, the convolutional neural network of artificial neurons is not influenced, during its training, by any variations in the dimensions of the handwriting samples within the same set. The modeling of the relevant features of the dermatoglyphs represented by the handwriting samples is then more accurate. The dimensions of the handwriting samples between the sets may nevertheless differ, particularly when the method is adapted to take handwriting samples of any size as input data, in accordance with certain embodiments described above.

[0066] According to some embodiments, the convolutional neural network CNN is trained on sets of training fingerprints which have been previously augmented by fingerprints selected from said sets at one or more predefined frequencies, said selected fingerprints having previously been subjected to an amputation and / or transformation operation by axial symmetry.

[0067] Adding amputated handwriting samples to training sets is particularly advantageous because it reduces the method's sensitivity to the potential absence of certain elements of the handwriting samples, for example, the absence of one or more fingers in the case of a hand that has undergone surgical amputation. The handwriting sample amputation operation can, for example, consist of the deliberate removal of certain fingers or regions of a handwriting sample selected from a training set. This can be performed at a frequency of 1 / 6, meaning that within the training set, one handwriting sample out of every six is ​​selected and undergoes amputation. The resulting new handwriting sample is then added to the training set.

[0068] Adding finger-sketches that have undergone axial symmetry is advantageous because it improves the method's sensitivity to chirality and harmonizes the distribution of finger-sketches between right-handed and left-handed characters within a training set. The axial symmetry operation can, for example, consist of performing an axial symmetry operation on a finger-sketch selected from a training set, and the resulting new finger-sketch is then added to the set, its class having been previously modified according to whether it represents the left or right hand. This operation can be performed at a frequency of 1 / 2, meaning that within the training set, every other finger-sketch is selected and axially symmetriced.

[0069] Training sets of typescripts can also be augmented using other known methods such as scaling changes, contrast changes, rotation operations and padding operations.

[0070] The method according to the first aspect of the invention is implemented by computer. With reference to [Fig. 5], in a second aspect of the invention, a data processing device 5000 is provided, comprising means for implementing a method 3000 according to any one of the embodiments of the first aspect of the invention.

[0071] An example of a device may be a device responsible for automatically executing sequences of arithmetic or logical operations to perform tasks or actions. This device, also called a computer, may include one or more central processing units (CPUs) and / or one or more graphics processing units (GPUs) 5001, as well as at least one control device suitable for performing these operations. It may also include other electronic components such as input / output interfaces 5002, non-volatile or volatile storage devices 5003, and communication buses for transferring data between internal components of the device or with external components. One of the input / output devices 5002 may be a user interface for human-computer interaction, for example, a graphical user interface for displaying human-understandable information.

[0072] According to a third aspect of the invention, a computer program 15003 is provided comprising instructions which, when the program is executed by a computer, cause the computer to implement a method (3000) according to any one of the embodiments of the first aspect of the invention.

[0073] Any type of programming language, compiled or interpreted, can be used to implement the steps of the method of the invention. The computer program may be part of a software solution, that is to say, a collection of executable instructions, code, scripts or other elements, and / or databases.

[0074] According to a fourth aspect of the invention, a computer-readable recording medium 5003 is provided comprising instructions which, when executed by a computer, cause the computer to implement a method according to any one of the embodiments of the first aspect of the invention.

[0075] The computer-readable recording medium 5003 is preferably non-volatile memory, for example a hard disk drive or a solid-state drive. It may be removable storage media or non-removable storage media that is part of a computer.

[0076] The computer-readable recording medium 5003 can also be volatile memory within a removable medium. This can facilitate the deployment of the invention at numerous production sites.

[0077] The computer-readable recording medium 5003 may be part of a computer used as a server from which executable instructions may be downloaded and, when executed by a computer, cause the computer to execute a method according to one of the embodiments described in this document.

[0078] The computer program 15003 and the storage medium 5003 on which it is recorded can be implemented in a distributed computing environment, for example, cloud computing. The instructions can be executed on a server to which one or more client computers can connect and provide encoded data. as input data for a method according to any one of the embodiments of the first aspect of the invention. Once the data has been processed, the result can be downloaded and decoded on the client computer or sent directly, for example, in the form of instructions.

[0079] In a fifth aspect of the invention, a system for classifying typescripts into a plurality of membership classes is provided, said system comprising: - a device (1000) for acquiring typing data; - a data processing device (5000) according to the second aspect of the invention, said data processing device (5000) is further configured to receive and process D-typescripts acquired by the D-typescript acquisition device (1000).

[0080] According to certain embodiments, the data processing device (5000) is further configured to emit a warning signal when a typescript acquired by the acquisition device does not conform to the expected class for said typescript and / or is assigned to a non-conforming class by a system operator. In particular, the device can be configured according to any one of the three implementation examples of the method according to the first aspect of the invention described above. Example

[0081] To illustrate the performance of the method according to the invention, an example of the method is provided in accordance with certain embodiments described above. The implemented convolutional neural network is that described in relation to [Fig. 4]. It was trained on a training set comprising more than 3000 typescripts classified according to the 11 membership classes such as those illustrated in [Fig. 2].

[0082] Figure 6 shows the occurrences, expressed as percentages, of correct (broken line) and incorrect (solid line) classification of the typescripts as a function of the membership probability values ​​P_K inferred by the method according to the example. The graph shows that 95% of the typescripts that were correctly classified were classified with a probability greater than 0.9. And only 15% of the typescripts that were incorrectly classified were classified with a probability greater than 0.9.

[0083] By setting a reliability threshold of 0.9 for the probability of belonging, below which the prediction is considered unreliable, 5.4% of the fingerprints in the training set can be eliminated, and the percentage of correctly classified fingerprints out of the remaining 94.6% reaches 99.7%. It was found that the 5.4% of fingerprints in the training set thus eliminated corresponded to so-called difficult cases for which assigning a class of belonging is not easy even when subjected to the expertise of a human operator.

[0084] In use, the inference times for a typescript supplied to a method according to this example are on the order of milliseconds for a 1600-byte convolutional network with 400,000 parameters stored on 32 bits running on an Intel® Core™ i7-7700 CPU. References Literature patent

[0085] US 5572597 Al, LORAL CORP [US], 05.11.1996.

[0086] EP 0 779 595 A2, NEC CORP [JP], 06.18.1997.

[0087] US 5,825,907 Al, LUCENT TECHNOLOGIES INC [US], 10.20.1998.

[0088] US 2012 014569 Al, IB KOREA LTD [KR], 01.19.2012.

[0089] US 9 530 042 Bl, UNIV KING SAUD [SA], 12 / 27 / 2016.

[0090] US 2017 046554 Al, NEC CORP [JP], 02 / 16 / 2017.

[0091] EP 3 825 915 Al, IDEMIA IDENTITY & SECURITY FRANCE [FR], 26.05.2021. Littérature non-brevet

[0092] F. Galton, Fingerprint Directories. London, MacMillan & Co, 1895.

[0093] Henry Faulds, Guide to fingerprint Identification, Tokyo, Hanley, 1905.

[0094] E. Henry, Classification and uses of finger prints, published by his majesty’s stationery office, London, 1913.

[0095] Wang et al., Fingerprint Classification Based on Depth Neural Network, arXiv preprint arXiv: 1409.5188, 2014.

[0096] Michelsanti et al., Fast Fingerprint Classification with Deep Neural Networks, VISIGRAPP 2017.

Claims

1. Demands Method (3000), implemented by computer, for classifying fingerprints into a plurality of membership classes Cj, each of which corresponds to a particular anatomical area Z_i of the palmar surface of a hand, said method (2000) takes, as input data (13000), at least one fingerprint D, and provides, as output data (03000), a membership class C or a list L of membership classes C_k of the fingerprint D from among the plurality of membership classes Cj, said method (3000) comprises the following steps: - provide (3001) a convolutional neural network (CNN) of artificial neurons, said convolutional neural network being previously trained on a training dataset E consisting of a plurality of handwriting samples D_i classified according to a plurality of classes CJ, each of which corresponds to a particular anatomical area of ​​a hand, said convolutional neural network being configured to provide the membership probabilities Pj of each handwriting sample D_i of the set E to each of the classes of the plurality of membership classes Cj; - infer (3002), using the trained convolutional neural network CNN, the membership probabilities P_i of the typescript D provided as input data (13000) to each of the classes of the plurality of membership classes Cj; - select (3003) a membership class C from among the membership classes Cj for which the membership probability P of the fingerprint D provided as input data (13000) is the highest among the membership probabilities P_i inferred for said classes C_j, or a list L of a number k of membership classes C_k selected from the plurality of membership classes Cj and sorted in ascending or descending order of the probabilities P_k of membership in said classes C_k inferred for the fingerprint D provided as input data (13000); the classes of the plurality of membership classes are selected from the complete right hand (C1), the complete left hand (C2), a lower area of ​​the palm of the right hand (C3), a lower area of ​​the palm of the left hand (C4), a upper area of ​​the right hand (C5), an upper area of ​​the left hand (C6), the writer's right palm (C7), the writer's left palm (C8), at least two, preferably three, preferably four fingers of the right hand (C9), at least two, preferably three, preferably four fingers of the left hand (CIO), the thumbs of the right and left hands (CIL).

2. Method (3000) according to claim 1, wherein the convolutional CNN of artificial neurons comprises: - a first sequence S1 of convolutional layers of artificial neurons; - a second sequence S2 of residual convolutional blocks of artificial neurons; - a third sequence S3 comprising at least one layer of fully connected artificial neurons.

3. Method (3000) according to claim 2, wherein the convolutional CNN of artificial neurons further comprises, between the first sequence SI and the second sequence S2, a spatial resampling step, preferably by bilinear interpolation.

4. Method (3000) according to any one of claims 1 to 3, wherein the cost function used in training the convolutional neural network CNN includes a cross-entropy type cost function.

5. Method (3000) according to any one of claims 1 to 4, wherein the cost function used in training the convolutional neural network (CNN) of artificial neurons comprises an axial symmetry sensitivity function, said axial symmetry sensitivity function being implemented in the following steps: (a) create a symmetrical fingerprint by an axial symmetry operation for each fingerprint in a training set; (b) calculate the membership probabilities of each fingerprint and its symmetrical counterpart to each of the classes in the plurality of membership classes as training input data by providing them to the convolutional neural network; (c) assign the membership probability values ​​of a class corresponding to an anatomical area of ​​one hand to the corresponding class of the other hand; (d) Integrate, into the cost function, a sensitivity function in the form of a Euclidean distance to be minimized between the membership values ​​assigned to the classes during step (c).

6. Method (3000) according to any one of claims 1 to 5, such that the convolutional neural network CNN is trained using a plurality of sets of training typescripts, each set comprising typescripts of identical dimension.

7. Method (3000) according to claim 6, wherein the convolutional neural network CNN is trained on sets of training handwriting samples having been previously augmented by handwriting samples selected from said sets at one or more predefined frequencies, said selected handwriting samples having been subjected to an amputation and / or axial symmetry transformation operation.

8. Data processing device (5000) comprising means for implementing a method (3000) according to any one of claims 1 to 7.

9. Computer program (15003) comprising instructions which, when the program is executed by a computer, cause the computer to implement a method (3000) according to any one of claims 1 to 7.

10. Computer-readable medium (5003) comprising instructions which, when executed by a computer, cause the computer to implement a method (3000) according to any one of claims 1 to 7.

11. A fingerprint classification system among a plurality of classes Cj of membership, each of the classes corresponding to a particular anatomical area Z_i of the palmar surface of a hand, said system comprises: - a fingerprint acquisition device (1000); - a data processing device (5000) according to claim 8, said data processing device (5000) is further configured to receive and process fingerprints acquired by the fingerprint acquisition device (1000).

12. A system according to claim 11, wherein the data processing device (5000) is further configured to emit a warning signal when a typescript is acquired by the device

13. acquisition does not conform to the expected class for said typescript and / or is assigned to a non-conforming class by a system operator. Use of a method according to any one of claims 1 to 7 to verify the integrity of a database of typescripts of different classes.

Images