Method for determining corrections to statistical discrepancies in the identification of a biometric identification system

Abstract

Method (500) for determining corrections to statistical discrepancies in the identification of a biometric identification system (200) by facial and / or pedestrian recognition, said method (500) comprises the following steps: (a) Define (501) a set of categories (C1-Cn), each category corresponding to at least one physical appearance characteristic; (b) Extract (502), for each individual (101) in the plurality of individuals, at least one physical appearance characteristic from at least one image of said individual (101); (c) Assign (503 / 600), to each individual (101), at least one category (C1-Cn) according to at least one criterion relative (az) to the extracted physical appearance characteristic; (d) Distribute (504 / 600), for each category (Cp), each individual (101) in the plurality of individuals into two groups (E, R) according to the failure (E) or success (R) of their identification by the biometric identification system (200);(e) Calculate (505 / 700), for each category (Cp), a number (N[Cp]) of individuals to be selected from the group (R) corresponding to the success (R) of the identification so that the relative proportions of individuals in the group (E) corresponding to the failure (E) of the identification are substantially equal between all the criteria (az) relating to said category (Cp).;

Description

Title of the invention: Method for determining corrections to statistical discrepancies in the identification of a biometric identification system technical field

[0001] The present invention relates to a method and device for determining corrections to statistical discrepancies in the identification of a biometric identification system. It also relates to a biometric identification system comprising such a device. Technical background

[0002] It is common to verify the identity of individuals using identification protocols based on the comparison of certain biometric characteristics. The use of these protocols generally requires a preliminary enrollment step whereby an individual registers with an entity with which they share a certain amount of biometric information and information relating to their identity.

[0003] Both enrollment and identification rely on a step of acquiring the individual's biometric characteristics. For this purpose, the individual presents themselves in front of an acquisition device of an acquisition system, which acquires an image of an area of ​​interest from which certain relevant anthropomorphic and / or anthropometric characteristics can be extracted for subsequent biometric analysis by a biometric processing unit. The identification processes are generally automated to allow, in particular, walk-through identification, without the individual needing to present themselves in front of the acquisition device. Restricted access areas and border control zones are examples of locations where automated walk-through identification systems are implemented. Summary of the invention

[0004] It has been observed that the algorithmic methods used in identification protocols can be affected by biases that lead to a number of identification failures. Human operator intervention is then necessary to complete, verify, or supplement the identification. These biases stem from incorrect detection or erroneous analysis of biometric information based on certain physical appearance characteristics of the individual. These physical appearance characteristics may include, in particular, innate physical characteristics, for example, physiological physical characteristics. and / or ethnic characteristics such as age, gender, skin color, or adventitious physical appearance features, for example, added elements, possibly in the form of bodily transformations, such as eyeglasses, tattoos, piercings, clothing worn on or around the head.

[0005] The origins of detection or analysis errors are diverse. They may result from an underrepresentation of certain groups of individuals in the training datasets of the algorithms, or from optical effects related to the surrounding conditions of the acquisition device of the identification system such as, for example, inappropriate lighting causing reflections on reflective surfaces such as the lenses of eyeglasses.

[0006] A major drawback of these biases is the discrimination against certain groups of individuals based on their physical appearance, particularly their physical and / or ethnic characteristics. This results in unfair treatment of these individuals during identity and / or access checks, as they are likely to be subject to more frequent checks by a human operator.

[0007] According to a first aspect of the invention, a method is provided, implemented by a data processing device, for determining corrections to statistical discrepancies in the identification of a biometric identification system using facial and / or pedestrian recognition. The method takes as input data a set of images comprising at least one image of each individual from a plurality of individuals acquired by said biometric identification system and the set of success or failure states of identification of each individual by said biometric identification system. The method provides as output data a set of numbers of individuals to be selected from each category of a set of categories relating, each, to at least one physical appearance characteristic. The method comprises the following steps: (a) Define a set of categories, each category corresponding to at least one physical appearance characteristic; (b) Extract, for each individual in the plurality of individuals, at least one physical appearance characteristic from at least one image of said individual; (c) Assign, to each individual, at least one category based on at least one criterion relating to the extracted physical appearance characteristic; (d) For each category, divide each individual from the plurality of individuals into two groups according to the success or failure of their identification by the biometric identification system; (e) Calculate, for each category, a number of individuals to be selected from the group corresponding to successful identification such that the relative proportions of individuals in the group corresponding to failure the identification should be substantially equal across all criteria relating to said category.

[0008] According to some embodiments, the process further comprises, before step a, an extraction step, from the plurality of individuals, of a sample of individuals comprising individuals representative of the physical appearance characteristics to which at least one category corresponds, steps b to e then being applied to said sample of individuals.

[0009] According to certain embodiments, the image set comprising at least one image of each individual of a plurality of individuals acquired by said biometric identification system and the set of success or failure states of identification of each individual by said biometric identification system are a sample of the identification history of said biometric identification system over a fixed period of time.

[0010] According to some embodiments, step b of extracting at least one physical appearance characteristic and step c of assigning at least one category are carried out using a convolutional neural network previously trained.

[0011] According to some embodiments, the physical appearance characteristic is chosen from innate physical appearance characteristics and / or adventitious physical appearance characteristics.

[0012] 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.

[0013] In a third aspect of the invention, a computer program is provided comprising instructions which, when the program is executed by a data processing device, cause the latter to implement a process according to the first aspect of the invention.

[0014] In a fourth aspect of the invention, a data processing device readable storage medium is provided comprising instructions which, when executed by a data processing device, cause the latter to implement a process according to the first aspect of the invention.

[0015] In a fifth aspect of the invention, a biometric identification system is provided comprising a data processing device according to the second aspect of the invention.

[0016] In a sixth aspect of the invention, a method for correcting statistical discrepancies in the identification of a biometric identification system using facial and / or pedestrian recognition is provided, said method comprising the following steps: (a) Determine the corrections to the statistical discrepancies in the identification of the biometric identification system using a method according to the first aspect of the invention; (b) Select, preferably randomly, for each category individuals in the group corresponding to success in identification; (c) Control the selected individuals by a human agent.

[0017] In a seventh aspect of the invention, a use of a method according to the first aspect of the invention is provided for correcting statistical discrepancies in the identification of a biometric identification system by facial and / or pedestrian recognition in a border control area. Brief description of the drawings

[0018] [Fig-1] is a schematic representation of an example of a border control area including a free-flow biometric identification sub-area.

[0019] [Fig.2] is a schematic representation of an example of an identification system biometric.

[0020] [Fig.3] is a schematic representation of an example of a processing device biometric.

[0021] [Fig.4] is an example of a graphical representation of success frequencies and failure to identify, by a biometric identification system, a plurality of individuals according to several categories of physical appearance characteristics.

[0022] [Fig.5] is a flowchart of a process according to an embodiment of the invention.

[0023] [Fig.6] is a schematic graphical representation of the success frequencies and failure to identify, by a biometric identification system, a plurality of individuals according to several categories of physical appearance characteristics, before correction.

[0024] [Fig.7] is a graphical representation of the frequencies of success and failure identification of [Fig. 6] after correction using a method according to the invention. Detailed description of embodiments

[0025] In this disclosure, embodiments are described in the general context of one or more hardware or devices capable of executing preloaded instructions such as, for example, computer-executable instructions for executing program modules. Program modules may include one or more routines, programs, objects, variables, commands, scripts, functions, applications, components, or data structures that can perform specific tasks or implement specific types of abstract data.

[0026] Certain embodiments can also be implemented in distributed computing environments where tasks are performed by remote data processing devices connected by a communication network. In a distributed computing environment, program modules may reside on local and / or remote computer storage media, including memory storage devices.

[0027] With reference to [Fig.1], a border control zone 100, such as an airport is likely to have, travellers 101 generally have the option of choosing between a sub-zone 102 of consent to biometric identification and a sub-zone 103 of non-consent to biometric identification, to access the gates 104 of boarding, when leaving the territory, or of disembarking, when arriving in the territory.

[0028] The biometric identification non-consent sub-zone 103 consists of a simple passageway in which travelers 101 wait before approaching a counter 105 where a human agent 106 is responsible for verifying their identity. In accordance with applicable national legislation, and respecting travelers' wishes that their anthropomorphic and / or anthropometric characteristics not be subject to biometric analysis, the non-consent sub-zone 103 may be devoid of acquisition and / or recording devices capable of providing evidence of these characteristics, or if it is equipped with such devices, these devices may not be configured to transmit this evidence to a biometric analysis system.

[0029] The biometric identification consent sub-zone 102 is, however, equipped with one or more devices 107, 108 for acquiring and / or recording tests of the anthropomorphic and / or anthropometric characteristics of travelers 101 in order to analyze them. In the example of a control zone shown in [Fig. 1], the acquisition devices 107, 108 are two aerial monocular cameras positioned on either side of a first room 102a that travelers 101 are asked to pass through. The two cameras 107, 108 are oriented at a downward angle in order to acquire one or more tests of the faces of travelers 101 who have previously completed their enrollment, for example, during a prior registration step (“check-in”).The tests are then transmitted to a biometric processing device (not shown) which extracts biometric test templates to be compared to reference biometric templates in a database and to identify travelers based on this comparison.

[0030] After passing through the first room 102a, the travelers enter a second room 102b adjoining the first, in which the unidentified travelers 101 Passengers using the identification system are directed to the control window 105 by a human control agent 109 for manual identification. Simultaneously, correctly identified passengers 101 are allowed access to the boarding or disembarking gates 104. The second room 102b may include a control system consisting, in this example, of two surveillance cameras 110, 111, positioned on either side of the room, to prevent any attempt by an unidentified passenger to access the boarding or disembarking gates 104.

[0031] With reference to [Fig. 2], a biometric identification system 200 that can be used for implementing free-flow identification in a border control zone 100 may be a facial and / or pedestrian recognition system. The system 200 comprises a biometric processing device 201 and one or more videographic and / or photographic recording devices 202, 203, such as one or more monocular cameras. For security reasons, the biometric processing device 201 is located away from the videographic and / or photographic recording devices 202, 203, generally in a dedicated room. It communicates with the videographic recording devices 202, 203 by any suitable wired or electromagnetic telecommunication device.

[0032] The recording device(s) 202, 203 are configured to record an image of an area 204a of interest of an individual 204, from which certain relevant anthropomorphic and / or anthropometric characteristics of the individual 204 can be extracted and converted into a proof biometric template by the biometric processing device 201. The biometric processing device 201 then compares the proof biometric template with one or more reference biometric templates from a database. If there is a match between the proof biometric template and at least one reference biometric template associated with an identified individual in the database, the individual 104 is considered identified by the identification system 200. Otherwise, the individual 104 is not considered identified. It must then be subject to manual control by a human agent 106.

[0033] The area 204a of interest of the individual 204 depends on the type of recognition implemented by the biometric identification system 200. In the case of facial recognition, the area 204a of interest includes the face and / or one or more characteristic facial features, such as the eyes, the corners of the mouth, or the nose. In the case of pedestrian recognition, the area 204a of interest includes at least the upper part of the individual 204's body, or even their entire body, and the biometric processing device 201 extracts a biometric template based on an analysis of the individual's posture, body size and / or gait 104.

[0034] Examples of biometric identification systems and methods using facial recognition are described in; Al Xinyi, et al. "A survey of face recognition." arXiv preprint arXiv:2212.13038 (2022); US 2017 / 0330028 Al [Idemia Identity and Security USA LLC] 16.11.2017; EP 3 285 209 A2 [Safran Identity and Security SAS] 21.02.2018.

[0035] Examples of biometric identification systems and methods using facial recognition are described by pedestrian recognition in Ye, Mang, et al. "Deep leaming for person re-identification: A survey and outlook." IEEE transactions on pattern analysis and machine intelligence 44.6 (2021): 2872-2893; US 2017 / 0316255 Al [Panasonic Intellectual Property Management Co Ltd Wang] 02.11.2017; WO2019 / 188111 Al [NEC CORP] 03.10.2019; US 2015 / 0193686 [Tata Consultancy Services Ltd] 09.07.2015.

[0036] With reference to [Fig. 3], a biometric processing device 201 is generally a data processing device 300 comprising means for implementing a biometric identification method as described above. This device 300 comprises one or more central processing units (CPUs) 301 and / or one or more graphics processing units (GPUs) 302, a physical remote communication module 303, one or more physical input / output modules 304 for exchanging data with external devices, a transient storage medium 305 such as random access memory (RAM), a non-transient recording medium 306, and communication buses (not shown) for transferring data between the internal components of the device 300.It may also include a secure 308 element for storing cryptographic keys, executing encryption algorithms, and / or storing and / or encrypting any other algorithm and / or data whose security and confidentiality must be preserved, for example, a database of reference biometric templates.

[0037] The biometric processing device 203 allows the execution of one or more program modules comprising instructions which, when the program module(s) are executed, cause the biometric processing device 203 to implement a biometric identification process as described above. The program module(s) may be written in any programming language, compiled or interpreted. They may be part of a software solution, i.e., a collection of executable instructions, code, scripts, or other components, and / or databases.

[0038] As explained previously, the algorithmic methods used in biometric identification protocols can be affected by biases that lead to a number of identification failures. These biases stem from incorrect detection or erroneous analysis of biometric information based on certain innate and / or adventitious physical appearance characteristics of the individual. This results in discrimination against certain groups of individuals based on their physical appearance characteristics, particularly their physical and / or ethnic physiological characteristics. These groups are then more frequently subject to unfair treatment during automated identity checks, requiring a human operator to intervene to compensate for the failures of automated identification and to perform the verification manually.

[0039] With reference to [Fig. 4], one or more categories C1-C6 relating to one or more physical appearance characteristics can be assigned to each individual in a group 401 of travelers 401a-d. Each category, discrete or continuous, corresponds to the application of a criterion relating to a physical appearance characteristic. "Physical appearance characteristic" means any innate or adventitious physical appearance characteristic of a group of individuals that allows for a distinction between the individuals of that group based on at least one objective criterion specific to that characteristic. Examples of innate physical appearance characteristics are physiological, physical, and / or ethnic body characteristics such as age, height, gender, skin color, facial shape, or eye color or shape.Examples of adventitious physical appearance features include reported elements, possibly in the form of bodily transformations, such as eyeglasses, tattoos, piercings, or clothing worn on or around the head.

[0040] In the example of [Fig.4], the first category Cl corresponds to the wearing Cia or not Clb of eyeglasses, the second category C2 comprises three intervals C2a, C2b, C3c of skin tone values, the third category C3 corresponds to the presence C3a or not C3b of a tattoo on the face, the fourth category C4 comprises five intervals C4a, C4b, C4c, C4d, C4e of age values, the fifth category C5 corresponds to the wearing C5a or not C5b of a head covering such as a hat, veil or scarf, and the sixth category C6 corresponds to the biological gender male C6a and the biological gender female C6b.

[0041] Each category C1-C6 corresponds to the application of a criterion relating to a physical appearance characteristic: the first category C1 corresponds to the criterion of wearing C1a or not wearing C1b eyeglasses; the second category C2 corresponds to the application of a criterion C2a, C2b, C3c of skin tone values; the third category C3 corresponds to the application of a criterion of presence C3a or non-presence C3b of facial tattoo; the fourth category C4 includes the application of a criterion C4a, C4b, C4c, C4d, C4e of age values; the fifth category C5 corresponds to the application of a criterion of wearing C5a or not wearing C5b of a head covering; and the sixth category C6 corresponds to the application of a criterion of male biological sex C6a or female biological sex C6b.

[0042] It should be noted that, generally, for any category associated with a physical appearance characteristic that can be measured on a continuous scale, it is possible, instead of discrete subcategories, to use a continuous scale of values ​​with a thresholding function. For example, for the third C3 and fifth C5 categories, instead of intervals, a continuous scale with thresholds can be used for the skin tone value and the age value, respectively.

[0043] Each traveler 401a-e in group 401 can be assigned one or more categories C1-C6 by applying at least one distinguishing criterion specific to the physical appearance characteristic associated with each of the categories C1-C6. For example, assuming that traveler 401a is a man in his thirties, wearing glasses, with a dark complexion and no tattoos, the categories C1a, C2c, C3a, C4c, C5b, and C6a are assigned to him. In [Fig. 4], the distribution of travelers 401 into the different categories C1-C6 is represented as a histogram of the absolute proportions of individuals in each category.

[0044] When a biometric identification algorithm is affected by bias, it may fail to identify more individuals from one or more categories than from others. Individuals can thus be divided into two groups E and R depending on whether their identification by the biometric identification algorithm was successful (E) or failed (R).

[0045] For example, because he wears glasses and / or has a dark complexion, traveler 401a may not be identified by the biometric identification system. He will then have to undergo manual identification by an operator. Any other traveler with the same characteristics may face a similar situation. This constitutes a form of discrimination against these individuals since they are more frequently subject to manual checks.

[0046] By generalizing, Figure 4, purely illustratively, shows the relative proportions of failure E (black part of the histogram) and success R (white part of the histogram) of identification for each histogram of proportions of individuals in each category C1-C6. It appears that travelers falling into categories C1, C2, C4, and C5 according to the respective criteria C1a, C2c, C4a, and C5a are not statistically identified more frequently. The behavior of the biometric identification system implementing an identification algorithm Biometric bias can then be considered discriminatory with regard to the physical appearance characteristics associated with these categories.

[0047] The objective of the present invention is to reduce, or even eliminate, identification biases that may affect current or future biometric identification algorithms. Another objective is to provide a solution adaptable to each situation in which these algorithms are implemented, thereby reducing the need to replace them, in whole or in part.

[0048] To this end, with reference to [Fig. 5] & [Fig. 6] & [Fig. 7], a method 500 is provided, implemented by a data processing device 300, for determining corrections to statistical discrepancies in the identification of a biometric identification system 200 using facial and / or pedestrian recognition. The method 500 takes as input a set 1501 of images Im comprising at least one image of each individual 101 of a plurality of individuals acquired by said biometric identification system 200 and the set 1502 of success states R or failure states E of identification of each individual 101 by said biometric identification system 200. The method 500 provides as output a set 0500 of numbers N[Cp] of individuals to be selected from each category Cp of a set of categories Cl-Cn, each relating to the minus one physical appearance characteristic, said process 500 comprises the following steps: (a) Define 501 a set of Cl-Cn categories, with each category corresponding to at least one physical appearance characteristic; (b) Extract 502, for each individual 101 of the plurality of individuals, at least one physical appearance characteristic from at least one image of said individual 101 (c) Assign 503 / 600, to each individual 101, at least one Cl-Cn category based on at least one criterion related to the extracted physical appearance characteristic; (d) Distribute 504 / 600, for each category Cp, each individual 101 of the plurality of individuals into two groups E, R according to the failure E or success R of their identification by the biometric identification system 200; (e) Calculate 505 / 700, for each category Cp, a number N[Cp] of individuals to be selected from the group R corresponding to the success R of the identification so that the relative proportions of individuals in the group E corresponding to the failure E of the identification are substantially equal between all the criteria a-z relative to said category Cp.

[0049] The method 500 according to the invention provides, for each category Cp, a number N[Cp] of individuals to be checked in the group R corresponding to the success R of the identification. In other words, for each category Cp, a number N[Cp] Individuals correctly identified (group R) by the biometric identification system 200 will be considered unidentified (group E) and will be subject to manual identification by a human agent 106. Thus, during a future identification campaign by the biometric identification system 200, for each category Cp, the relative proportions of individuals in group R will be substantially equal between all the criteria az relating to said category Cp, so as, for example, to obtain, for each category Cp, a distribution between successes R and failures E as represented on the histogram of [Fig.7].

[0050] The set 0500 of N[Cp] numbers calculated for each Cp can, for example, be transmitted to the biometric identification system 200, which will then proceed to select, preferably randomly, for each category Cp, the N[Cp] individuals in group R for manual identification by a human agent. With reference to [Fig. 1], in the example of a border control zone 101, the N[Cp] travelers 101 selected by the identification system 200 are redirected by a control agent 109 to the window 105 where a human agent 106 is responsible for verifying their identity. For this purpose, the control agent 109 may be equipped with a mobile electronic device 112 on which he is notified of the travellers 101 who must be manually identified, including both those selected by the biometric identification system 200 and those whom the biometric identification system 200 has not been able to identify.Preferably, neither the control officer 109 nor the identification officer 106 should be aware of the reason, namely selection or non-identification, for which these travelers 101 are being redirected to manual identification.

[0051] According to a purely illustrative example of the method according to the invention, a category Cl corresponding to biological gender is defined. After extracting the physical appearance characteristics relevant for gender identification, each individual 101 of a plurality of individuals is assigned to a biological gender category Cl based on two criteria a, b, corresponding respectively to the female biological gender Cia and the male biological gender Clb. The division of individuals into two groups according to the success (E) or failure (R) of their identification reveals, for example, that 4% of the individuals 101 of the male biological gender Clb are not identified and 2% of the individuals 101 of the female biological gender Cia are not identified.Assuming that the biometric identification system 200 identifies an average of 1000 individuals per day, comprising 300 individuals of the male biological sex Clb and 700 individuals of the female biological sex Cia, 12 individuals of the male biological sex Clb and 14 individuals of the female biological sex Cia are therefore not identified. Although the number of unidentified individuals between the two criteria is very close, the relative proportion of individuals... 101 of the male biological gender Clb is twice as high as that of individuals 101 of the female biological gender Cia.

[0052] In order to restore the balance between the relative proportions of identification failure between individuals of the male biological sex Cia and individuals of the female biological sex Clb, the calculation step 505 / 700 determines the number of people to be selected from the group corresponding to the success rate R of identification for each of the two criteria a, b, corresponding respectively to the female biological sex Cia and the male biological sex Clb. In the example, 39 individuals of the female biological sex Cia correctly identified and 11 individuals of the female biological sex Cia correctly identified will be selected, preferably randomly, during the next implementation of the biometric identification system.The relative proportions of unidentified individuals (101), or those considered as such, between the two criteria a and b, corresponding respectively to the female biological gender Cia and the male biological gender Clb, will then be essentially identical. In this case, 7.67% of individuals (101) of the female biological gender Cia and 7.67% of individuals (101) of the male biological gender Clb will not be identified or considered as such.

[0053] Depending on the type and performance of the biometric identification algorithm implemented by the biometric identification system 200, certain physical appearance characteristics may influence the outcome of the identification of individuals 101 to a greater or lesser extent. The identification, by said system 200, of individuals 101 belonging to categories relating to these physical appearance characteristics is therefore likely to fail more frequently than that of individuals in other categories. For example, for the reasons mentioned above, a biometric identification algorithm may fail more often to identify individuals wearing eyeglasses than individuals with different physical appearance characteristics. Or again, a biometric identification algorithm may have an acceptable identification failure rate for individuals belonging to certain categories but not for others.

[0054] Also, according to certain embodiments, the process further comprises, before step 502, a step 502a of extracting, from the plurality of individuals, a sample of individuals comprising individuals representative of the physical appearance characteristics to which at least one category Cp corresponds, steps 502 to 505 then being applied to said sample of individuals. Such sampling makes it possible to concentrate the corrections of statistical identification discrepancies on individuals for whom the biometric identification system 200 is most deficient, while preserving its identification performance for other individuals. Because process 500 limits the corrections to only a number limited in categories of individuals, it is more economical in computing resources and therefore in energy.

[0055] The method 500 according to the invention can be implemented in real time or with a delay relative to the operating period of the biometric identification system 200. According to certain embodiments, the set 1501 of images comprising at least one image of each individual 101 of a plurality of individuals acquired by said biometric identification system 200 and the set 1502 of the success R or failure E identification states of each individual 101 by said biometric identification system 200 constitute a sample of the identification history of said biometric identification system 200 over a fixed period of time.

[0056] In real-time implementation, the fixed time period can, for example, be a rolling period regularly updated at a given frequency to account for the most recent identification operations as the biometric identification system 200 is used. At each update, the method 500 according to the invention performs a new determination of the corrections to be applied to said system 200 and transmits to it an updated set 0500 of numbers N[Cp] of individuals calculated for each Cp that it will now have to select for manual identification. This update operation takes place at a given frequency, for example daily, throughout the entire operating time of the biometric identification system 200.

[0057] In a delayed implementation, a sample of the identification history of the biometric identification system 200 is provided as input data to the process 500 according to the invention at the end of an individual identification campaign by said system 500. From this sample, the process 500 according to the invention performs a new determination of the corrections to be applied to said system 200 and transmits to it an updated set 0500 of numbers N[Cp] of individuals calculated for each Cp. Then, during the next identification campaign, the system 200 will apply this update.

[0058] In step 502, for each individual 101 in the plurality of individuals, at least one physical appearance characteristic is extracted from at least one image of said individual using any suitable method. According to some preferred embodiments, step 502, extracting at least one physical appearance characteristic, and step 503, assigning at least one Cl-Cn category, are performed using a previously trained convolutional neural network. Examples of convolutional neural networks suitable for determining a person's gender and age are described in Levi, Gil, and Tal Hassner, "Age and gender classification using convolutional neural networks," Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2015; Kuprashevich, Maksim, and Irina Tolstykh. "Mivolo: Multi-input transformer for age and gender estimation." International Conference on Analysis of Images, Social Networks and Texts. Cham: Springer Nature Switzerland, 2023.

[0059] In a second aspect of the invention, with reference to [Fig.3], the method 500 according to the invention can be implemented by a data processing device 300.

[0060] In a third aspect, the method 500 according to the invention is in the form of a computer program or a computer program module comprising instructions which, when the program is executed by a data processing device 300, implement said method 500.

[0061] In a fourth aspect of the invention, the computer program or computer program module is stored in a non-transient recording medium 306 of a data processing device 300.

[0062] In a fifth aspect of the invention, a biometric identification system 200 is provided, comprising a data processing device 300 according to the second aspect of the invention. Preferably, the data processing device 300 is constituted by the biometric processing device of said system 200.

[0063] The method 500 according to the first aspect of the invention and / or the system 200 according to the fifth aspect of the invention can advantageously be used for the correction of statistical discrepancies in the identification of a biometric identification system 200 by facial and / or pedestrian recognition in a border control zone 100 such as, for example, described in the context of [Fig.1].

[0064] To this end, in a sixth aspect of the invention, the method 500 according to the first aspect of the invention can advantageously be used to implement a method for correcting statistical discrepancies in the identification of a biometric identification system 200 using facial and / or pedestrian recognition. Such a correction method comprises the following steps: (a) Determine the corrections to the statistical discrepancies in the identification of the biometric identification system 200 using a method 500 according to any one of the embodiments of the first aspect of the invention; (b) Select, preferably randomly, for each category Cp, N[Cp] individuals in the group R corresponding to success in identification; (c) Control the selected individuals by a human agent. References Literature patent

[0065] US 2015 / 0193686 [Tata Consultancy Services Ltd] 07 / 09 / 2015

[0066] US 2017 / 0316255 Al [Panasonic Intellectual Property Management Co Ltd Wang] 02.11.2017.

[0067] US 2017 / 0330028 Al [Idemia Identity and Security USA LLC] 16.11.2017.

[0068] EP 3 285 209 A2 [Safran Identity and Security SAS] 02.21.2018

[0069] WO 2019 / 188111 Al [NEC CORP] 03.10.2019. Non-patent literature

[0070] Levi, Gil, and Tal Hassner. "Age and gender classification using convolutional neural networks." Proceedings of the IEEE conférence on computer vision and pattern récognition workshops. 2015.

[0071] Ye, Mang, et al. "Deep learning for person re-identification: A survey and outlook." IEEE transactions on pattern analysis and machine intelligence 44.6 (2021): 2872-2893.

[0072] Wang, Xinyi, et al. "A survey of face récognition." arXiv preprint arXiv:2212.13038 (2022).

[0073] Kuprashevich, Maksim, and Irina Tolstykh. "Mivolo: Multi-input transformer for âge and gender estimation." International Conférence on Analysis of Images, Social Networks and Texts. Cham: Springer Nature Switzerland, 2023.

Claims

1.

2. Demands A method (500), implemented by a data processing device (300), for determining corrections to statistical discrepancies in the identification of a biometric identification system (200) using facial and / or pedestrian recognition, the method (500) takes as input data a set (1501) of images (Im) comprising at least one image of each individual (101) from a plurality of individuals acquired by said biometric identification system (200) and the set (1502) of success (R) or failure (E) identification states of each individual (101) by said biometric identification system (200), said method (500) provides as output data a set (0500) of numbers (N[Cp]) of individuals to be selected from each category (Cp) of a set of categories (Cl-Cn) relating, each, to at least one physical appearance characteristic, said Process (500) comprises the following steps: (a) Define (501) a set of categories (Cl-Cn), to each category corresponding at least one physical appearance characteristic; (b) Extract (502), for each individual (101) of the plurality of individuals, at least one physical appearance characteristic from at least one image of said individual (101); (c) Assign (503 / 600), to each individual (101), at least one category (Cl-Cn) according to at least one criterion relating (az) to the extracted physical appearance characteristic; (d) Distribute (504 / 600), for each category (Cp), each individual (101) of the plurality of individuals into two groups (E, R) according to the failure (E) or success (R) of their identification by the biometric identification system (200); (e) Calculate (505 / 700), for each category (Cp), a number (N[Cp]) of individuals to be selected from the group (R) corresponding to the success (R) of the identification so that the relative proportions of individuals in the group (E) corresponding to the failure (E) of the identification are substantially equal between all the criteria (az)relative to said category (Cp). A process according to claim 1, wherein the process further comprises, prior to step (502), an extraction step (502a) from the plurality of individuals, from a sample of individuals comprising individuals representative of the physical appearance characteristics to which corresponds at least one category (Cp), steps (502) to (505) then being applied to said sample of individuals.

3. A method according to any one of claims 1 to 2, wherein the set (1501) of images comprising at least one image of each individual (101) of a plurality of individuals acquired by said biometric identification system (200) and the set (1502) of success (R) or failure (E) identification states of each individual (101) by said biometric identification system (200) are a sample of the identification history of said biometric identification system (200) over a fixed period of time.

4. A method according to any one of claims 1 to 3, wherein step (502) of extracting at least one physical appearance feature and step (503) of assigning at least one category (Cl-Cn) are carried out using a previously trained convolutional neural network.

5. A method according to any one of claims 1 to 4, wherein the physical appearance characteristic is chosen from innate physical appearance characteristics and / or adventitious physical appearance characteristics.

6. Data processing device (300) comprising means for implementing a method (500) according to any one of claims 1 to 5.

7. A computer program comprising instructions which, when the program is executed by a data processing device (300), cause the device to implement a method (500) according to any one of claims 1 to 5.

8. Data processing device (306) readable storage medium (300) comprising instructions which, when executed by a data processing device (300), cause the device (300) to implement a method (500) according to any one of claims 1 to 5.

9. Biometric identification system (200) comprising a data processing device (300) according to claim 6.

10. A method for correcting statistical discrepancies in the identification of a biometric identification system (200) using facial and / or pedestrian recognition, said method comprising the following steps: (a) Determine the corrections for statistical discrepancies in the identification of the biometric identification system (200) using a method (500) according to any one of claims 1 to 5; (b) Select, preferably randomly, for each category (Cp), N[Cp] individuals in the group (R) corresponding to successful identification; (c) Control the selected individuals by a human agent.

11. Use of a method (500) according to any one of claims 1 to 5 for correcting statistical discrepancies in the identification of a biometric identification system (200) by facial and / or pedestrian recognition in a border control zone (100).

Images