Identity verification method and system

In voiceprint recognition and biometric identity verification, the probability distribution model and distribution parameters of different scenarios are used respectively, and the likelihood scoring is calculated in combination with mapping relationships, the recognition accuracy caused by scene mismatch and diversification in the existing technology is solved. The problem has been reduced, achieving higher recognition accuracy and stability.

CN112487384BActive Publication Date: 2024-12-03HUAWEI TECH CO LTD +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202011341050.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-11-25
Publication Date
2024-12-03
Estimated Expiration
2040-11-25

AI Technical Summary

Technical Problem

Existing voiceprint recognition technology and biometric authentication have robustness problems in registration and identification scenarios, resulting in a decrease in recognition accuracy and the inability to effectively deal with the causes of scene mismatch and diversification. The existing technology has not fully considered statistical significance. inconsistency and diversity causes.

Method used

By using different probability distribution models and distribution parameters in the registration and identification scenarios, compute likelihood scoring with mapping relationships to make up for the losses caused by scene mismatch, the mapping function is trained using the loss function estimated by linear Gaussian distribution and maximum likelihood criterion estimation. , use MLLR task to train mapping functions, dynamically update distribution parameters and clustering algorithms for scoring and fusion.

Benefits of technology

Effectively avoid the loss of identity authentication performance caused by scene mismatch, improve the recognition accuracy, and is suitable for biometric authentication such as voiceprint, face and fingerprint recognition, especially in cross-device and cross-channel scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112487384B_ABST
    Figure CN112487384B_ABST
Patent Text Reader

Abstract

The present application relates to the field of artificial intelligence and provides a method for identity authentication. The method comprises: obtaining a biometric feature extracted from a first registration object in a registration scenario; obtaining a biometric feature extracted from a first identification object in an identification scenario; calculating a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object; and judging whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score. The likelihood score includes a posterior probability part and a marginal probability part.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a method and system for identity authentication through biometric features of a user. Background Art

[0002] The user's biometric features can be used to register the user's biometric features, and the user's identity can be verified according to the registered biometric features during identification. For example, common face recognition, fingerprint recognition and voiceprint recognition can complete feature comparison and identity authentication by comparing the features left by the user in advance with the features collected during identification. Taking the common voiceprint recognition as an example, it is also called speaker recognition, which refers to the use of speaker feature information contained in the voice to automatically identify the identity of the interlocutor. Among them, voiceprint recognition usually includes two stages, the registration stage and the recognition stage. The registration stage refers to the speaker's reserved voice, extracting voiceprint features from the reserved voice, and then training the corresponding speaker model based on the voiceprint features and storing the model. The recognition stage refers to the recognition and authentication of the speaker, extracting voiceprint features from the recognized voice, and scoring the likelihood of similarity between the voiceprint features during recognition and the model stored in the model library, and finally determining the speaker's identity based on the recognition score. Voiceprint recognition based on deep neural networks usually uses a large amount of speaker training corpus to train deep neural networks, and then extracts voiceprint features from the original acoustic features through the trained neural network, and the extracted features represent the speaker characteristics. Commonly used voiceprint features include I-vector, TDNN-X-vector, RESNET-embedding, etc.

[0003] Voiceprint recognition technology and similar biometric-based identity authentication technologies face robustness issues during recognition. Taking voiceprint recognition as an example, in practice, it is often necessary to establish a user's voiceprint model based on a small amount of reserved voice information provided by the user. However, it is difficult for a small amount of reserved voice information to capture robust voiceprint features, resulting in the trained voice features being unable to cover the changes of the speaker himself, leading to mismatches during voiceprint registration and recognition. On the other hand, in the recognition scenario, the speaker's own voice characteristics may change over time and the state and manner of speaking may change. The recognition scenario may also have factors such as noise interference and channel equipment differences. There are also channel differences caused by different devices used for voiceprint registration and voiceprint recognition, which leads to changes in the speaker's feature space. These will also cause mismatches during voiceprint registration and recognition. The mismatch between the voiceprint registration scenario and the voiceprint recognition scenario will reduce the recognition accuracy and affect the user experience.

[0004] In the prior art, in order to cope with the mismatch between the voiceprint registration scenario and the voiceprint recognition scenario, the following methods are mainly used: enhance and transform the original data and features of the mismatched scenario to meet the registration scenario requirements as much as possible; optimize the training model to reduce the sensitivity of the user's voiceprint features to changes in the user or the scenario, thereby reducing the impact of the scene mismatch; model the scoring model of the specific scene when scoring at the back end (such as an adaptive PLDA model, etc.). However, the coping methods in the prior art are to weaken the impact of the scene mismatch on the change of voiceprint features in terms of data features or model training, and do not fully consider the objective statistical inconsistency of the data collected by the registration scenario and the recognition scenario, and therefore it is difficult to fundamentally make up for the differences caused by the scene mismatch. In addition, the methods of the prior art do not fully consider the diverse causes of scene mismatch. Therefore, a technical solution is needed that can accept the objective existence of scene mismatch and fully consider the diverse causes behind it. Summary of the invention

[0005] The purpose of this application is to provide a method and system for identity authentication in order to solve technical problems such as degradation of identity authentication performance caused by scenario mismatch in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a method for identity authentication. The method includes: obtaining a biometric feature extracted from a first registration object in a registration scenario; obtaining a biometric feature extracted from a first identification object in an identification scenario; calculating a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object; and judging whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score. Wherein, the likelihood score includes a posterior probability part and an edge probability part. The posterior probability part is determined based on the biometric feature of the first registration object, the probability distribution model of the registration scenario, and the mapping relationship between the probability distribution model of the registration scenario and the probability distribution model of the identification scenario. The probability distribution model of the registration scenario includes the probability distribution parameters between different registration objects and the probability distribution parameters between different samples of the same registration object when extracting biometric features from one or more registration objects in the registration scenario, and the probability distribution model of the identification scenario includes the probability distribution parameters between different identification objects and the probability distribution parameters between different samples of the same identification object when extracting biometric features from one or more identification objects in the identification scenario. The edge probability part is determined based on the biometric feature of the first identification object, the probability distribution model of the identification scenario, and the mapping relationship.

[0007] The technical solution described in the first aspect realizes the use of different probability distribution models and distribution parameters for registration scenarios and recognition scenarios respectively, and corrects the likelihood scores in the case of scene mismatch in combination with the mapping relationship, which is helpful to compensate for the losses caused by scene mismatch.

[0008] According to the first aspect, in a possible implementation method, the likelihood score is calculated based on the biometric characteristics of the first registered object and the biometric characteristics of the first identified object, including: according to the mapping relationship, mapping the biometric characteristics of the first identified object from the identification scene to the registration scene to obtain the mapped biometric characteristics of the first identified object; calculating the posterior probability distribution template based on the biometric characteristics of the first registered object on the probability distribution model of the registration scene; calculating the posterior probability part based on the posterior probability distribution template and the mapped biometric characteristics of the first identified object on the probability distribution model of the registration scene; calculating the marginal probability part based on the biometric characteristics of the first identified object on the probability distribution model of the identification scene; and obtaining the likelihood score based on the posterior probability part and the marginal probability part.

[0009] In this way, by mapping from the recognition scene to the registration scene, a corresponding likelihood scoring method is designed, which is helpful to avoid the loss of identity authentication performance due to scene mismatch.

[0010] According to the first aspect, in a possible implementation method, the likelihood score is calculated based on the biometric characteristics of the first registered object and the biometric characteristics of the first identified object, including: calculating a posterior probability distribution template on the probability distribution model of the registration scene based on the biometric characteristics of the first registered object, mapping the posterior probability distribution template from the registration scene to the identification scene according to the mapping relationship to obtain the mapped posterior probability distribution template, calculating the posterior probability part on the probability distribution model of the identification scene according to the mapped posterior probability distribution template, calculating the marginal probability part on the probability distribution model of the identification scene according to the characteristics of the first identified object, and obtaining the likelihood score based on the posterior probability part and the marginal probability part.

[0011] In this way, by mapping from the registration scene to the recognition scene, a corresponding likelihood scoring method is designed, which is helpful to avoid the loss of identity authentication performance due to scene mismatch.

[0012] According to the first aspect, in a possible implementation method, the probability distribution model of the registration scene and the probability distribution model of the recognition scene both obey linear Gaussian distribution, the probability distribution parameters between different registration objects include the mean and variance of the feature space distribution of different registration objects in the registration scene, the probability distribution parameters between different samples of the same registration object include the variance of the feature space distribution of the same registration object in the registration scene, the probability distribution parameters between different recognition objects include the mean and variance of the feature space distribution of different recognition objects in the recognition scene, and the probability distribution parameters between different samples of the same recognition object include the variance of the feature space distribution of the same recognition object in the recognition scene.

[0013] In this way, correction is achieved to compensate for the mismatch between different scenes that obey the linear Gaussian distribution.

[0014] According to the first aspect, in a possible implementation manner, the likelihood score is calculated according to the linear Gaussian distribution.

[0015] In this way, correction for mismatch between different scenes under Gaussian distribution is achieved.

[0016] According to the first aspect, in a possible implementation method, the mapping relationship is determined according to a trained mapping function, and the trained mapping function is obtained by the following method: estimating the feature distribution in the registration scenario according to the registration training data, estimating the feature distribution in the recognition scenario according to the recognition training data, and calculating a loss function based on the maximum likelihood criterion estimation according to the mapping function, the registration training data and the recognition training data, wherein the mapping function is used for the inverse mapping between the feature distribution in the registration scenario and the feature distribution in the recognition scenario, and adjusting the parameters of the mapping function according to the loss function until the loss function meets the minimization optimization condition, thereby obtaining the trained mapping function.

[0017] In this way, the deep learning algorithm is combined to train the mapping function and obtain the likelihood score.

[0018] According to the first aspect, in a possible implementation manner, the mapping function is a linear function of one variable, and the parameters of the mapping function include two constants of the linear function of one variable.

[0019] In this way, correction under mismatch between scenes satisfying the linear inverse mapping relationship is achieved.

[0020] According to the first aspect, in one possible implementation, the machine learning model corresponding to the loss function is used for a maximum likelihood linear regression (MLLR) task.

[0021] In this way, the mapping function is trained through the machine learning model of the MLLR task.

[0022] According to the first aspect, in a possible implementation, the method further includes: obtaining biometric features extracted from multiple samples of the second identification object in the identification scenario and calculating likelihood scores corresponding to each of the multiple samples of the second identification object, screening out features for state update from the biometric features extracted from the multiple samples of the second identification object according to the likelihood scores corresponding to each of the multiple samples of the second identification object, and updating the probability distribution parameters between different samples of the same identification object included in the probability distribution model of the identification scenario using the features for state update. In this way, the distribution parameters are adjusted and updated according to the biometric features of the multiple samples of the second identification object, which is conducive to dynamic adjustment and update according to changes in the identification object itself.

[0023] According to the first aspect, in a possible implementation, the method further includes: clustering the probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene through a clustering algorithm and obtaining multiple sub-models, wherein each of the multiple sub-models corresponds to a different state of the same recognition object; wherein the marginal probability part of the likelihood score fuses the score results of each of the multiple sub-models. In this way, the score fusion is achieved through the clustering algorithm, which is conducive to making adjustments according to the different states of the same recognition object.

[0024] According to the first aspect, in a possible implementation, the biometric feature includes a voiceprint feature, so as to correct the scenario mismatch of identity authentication based on voiceprint recognition.

[0025] According to the first aspect, in a possible implementation, the biometric feature includes a facial feature, so as to correct the scene mismatch of the identity authentication based on facial recognition.

[0026] According to the first aspect, in a possible implementation, the biometric feature includes a fingerprint feature, so as to correct the scenario mismatch of identity authentication based on fingerprint recognition.

[0027] According to the first aspect, in a possible implementation manner, the registration scenario and the identification scenario correspond to different devices, respectively, so that cross-device registration and identification are implemented.

[0028] In the second aspect, an embodiment of the present application provides a chip system, characterized in that the chip system is applied to an electronic device; the chip system includes one or more interface circuits and one or more processors; the interface circuit and the processor are interconnected through lines; the interface circuit is used to receive a signal from a memory of the electronic device and send the signal to the processor, the signal including computer instructions stored in the memory; when the processor executes the computer instructions, the electronic device executes the identity authentication method provided in the first aspect above.

[0029] The technical solution described in the second aspect realizes the use of different probability distribution models and distribution parameters for registration scenarios and recognition scenarios respectively through the chip system, and corrects the likelihood score in the case of scene mismatch in combination with the mapping relationship, which is helpful to compensate for the losses caused by scene mismatch.

[0030] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the processor executes the identity authentication method provided in the first aspect.

[0031] The technical solution described in the third aspect implements the use of different probability distribution models and distribution parameters for registration scenes and recognition scenes respectively through the computer-readable storage medium, and corrects the likelihood score in the case of scene mismatch in combination with the mapping relationship, thereby facilitating compensation for the losses caused by scene mismatch.

[0032] In a fourth aspect, an embodiment of the present application provides a computer program product, characterized in that the computer program product includes computer program instructions, and when the computer program instructions are executed by a processor, the processor executes the identity authentication method provided in the first aspect.

[0033] The technical solution described in the fourth aspect realizes the use of different probability distribution models and distribution parameters for registration scenes and recognition scenes respectively by executing the computer program instructions, and corrects the likelihood score in the case of scene mismatch in combination with the mapping relationship, thereby facilitating compensating for the losses caused by scene mismatch.

[0034] In a fifth aspect, an embodiment of the present application provides an identity authentication system. The identity authentication system includes: a registration module, which is used to obtain a biometric feature extracted from a first registration object in a registration scenario; an identification module, which is used to obtain a biometric feature extracted from a first identification object in an identification scenario; and a scoring judgment module, which is used to calculate a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object, and to judge whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score. Among them, the likelihood scoring includes a posterior probability part and a marginal probability part, the posterior probability part is determined according to the biometric characteristics of the first registered object, the probability distribution model of the registration scenario and the mapping relationship between the probability distribution model of the registration scenario and the probability distribution model of the recognition scenario, the probability distribution model of the registration scenario includes the probability distribution parameters between different registered objects and the probability distribution parameters between different samples of the same registered object when extracting biometric characteristics from one or more registered objects in the registration scenario, the probability distribution model of the recognition scenario includes the probability distribution parameters between different recognized objects and the probability distribution parameters between different samples of the same recognized object when extracting biometric characteristics from one or more recognized objects in the recognition scenario, and the marginal probability part is determined according to the biometric characteristics of the first recognized object, the probability distribution model of the recognition scenario and the mapping relationship.

[0035] The technical solution described in the fifth aspect realizes the use of different probability distribution models and distribution parameters for registration scenarios and recognition scenarios respectively, and corrects the likelihood scores in the case of scene mismatch in combination with the mapping relationship, which is helpful to compensate for the losses caused by scene mismatch. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to illustrate the technical solutions in the embodiments of the present application or the background technology, the drawings required for use in the embodiments of the present application or the background technology will be described below.

[0037] Figure 1 A flow chart of an identity authentication method according to an implementation method provided in an embodiment of the present application is shown.

[0038] Figure 2 A flow chart of another identity authentication method provided in an embodiment of the present application is shown.

[0039] Figure 3 A schematic diagram of identity authentication based on biometric registration and recognition provided in an embodiment of the present application is shown.

[0040] Figure 4 A structural block diagram of an identity authentication system provided in an embodiment of the present application is shown.

[0041] Figure 5A structural block diagram of an identity authentication system for dynamically matching user changes provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0042] The embodiment of the present application provides an identity authentication method and system. The identity authentication system includes: a registration module, which is used to obtain a biometric feature extracted from a first registration object in a registration scenario; an identification module, which is used to obtain a biometric feature extracted from a first identification object in an identification scenario; and a scoring judgment module, which is used to calculate a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object, and to judge whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score. Wherein, the likelihood score includes a posterior probability part and a marginal probability part, the posterior probability part is determined according to the biometrics of the first registered object, the probability distribution model of the registration scene, and the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene, the probability distribution model of the registration scene includes the probability distribution parameters between different registered objects and the probability distribution parameters between different samples of the same registered object when extracting biometrics from one or more registered objects in the registration scene, the probability distribution model of the recognition scene includes the probability distribution parameters between different recognized objects and the probability distribution parameters between different samples of the same recognized object when extracting biometrics from one or more recognized objects in the recognition scene, and the marginal probability part is determined according to the biometrics of the first recognized object, the probability distribution model of the recognition scene, and the mapping relationship. In this way, different probability distribution models and distribution parameters are used for the registration scene and the recognition scene respectively, and the likelihood score is corrected in the case of scene mismatch in combination with the mapping relationship, which is conducive to compensating for the loss caused by the scene mismatch.

[0043] The embodiments of the present application can be used in the following application scenarios: fields such as voiceprint recognition, face recognition, fingerprint recognition, etc. that require the use of biometric features for identity authentication, or application scenarios that need to deal with changes in the feature space of biometric features that lead to scenario mismatches, such as cross-device and cross-channel registration and identification.

[0044] The embodiments of the present application can be adjusted and improved according to the specific application environment and are not specifically limited here.

[0045] In order to enable those skilled in the art to better understand the present application, the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0046] See also Figure 1 , Figure 1 FIG. 1 is a flow chart showing an identity authentication method according to an embodiment of the present application. Figure 1As shown, the method includes the following steps.

[0047] Step S102: obtaining a biometric feature extracted from a first registration object in a registration scenario and a biometric feature extracted from a first identification object in an identification scenario.

[0048] It should be understood that the biometric feature can be a voiceprint feature, a facial feature, a fingerprint feature, or other biometric features used for identity authentication and identification. The definition of the first registration object and the first identification object is a person who provides a biometric feature and whose respective identities correspond to the provided biometric feature. For example, taking the biometric feature as a voiceprint feature as an example, the first registration object is the user who registers the voiceprint feature in the registration scenario, or can be called the speaker in the registration scenario. In contrast, the first identification object is the user who performs voiceprint feature identification in the identification scenario, or can be called the speaker in the identification scenario.

[0049] Step S104: Calculate a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object, wherein the likelihood score includes a posterior probability part and a marginal probability part.

[0050] Among them, the posterior probability part is determined according to the biometric feature of the first registration object, the probability distribution model of the registration scene and the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. The marginal probability part is determined according to the biometric feature of the first recognition object, the probability distribution model of the recognition scene and the mapping relationship. The probability distribution model of the registration scene includes the probability distribution parameters between different registration objects and the probability distribution parameters between different samples of the same registration object when extracting biometric features from one or more registration objects in the registration scene. The probability distribution model of the recognition scene includes the probability distribution parameters between different recognition objects and the probability distribution parameters between different samples of the same recognition object when extracting biometric features from one or more recognition objects in the recognition scene. The probability distribution model of the registration scene can be understood as the distribution law of the biometric features extracted in the limited space of the registration scene. Taking the biometric feature as the voiceprint feature as an example, the probability distribution model of the registration scene can be understood as the distribution of the voiceprint features of the speaker extracted in the limited space of the registration scene, or can be understood as a model of the feature distribution space used to describe the voiceprint features of the entire registration scene. In addition, there is a mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. This mapping relationship can be reflected by a mapping function, which can be linear or nonlinear. A simple linear mapping function is taken as an example below. This simple linear mapping function can be expressed as a linear function with two constants.

[0051] It is further assumed that the probability distribution model of the registration scenario and the probability distribution model of the recognition scenario both satisfy the Gaussian distribution law. This means that, taking the biometric feature as a voiceprint feature as an example, the distribution of the feature space of the speaker's voiceprint feature in the entire registration scenario conforms to the linear Gaussian distribution. Conversely, the distribution of the feature space of the speaker's voiceprint feature in the entire recognition scenario also conforms to the linear Gaussian distribution. It should be understood that the embodiments of the present application can also be applied to other types of distribution laws such as the Laplace distribution law. The following description is based on the premise that the probability distribution model of the registration scenario and the probability distribution model of the recognition scenario both satisfy the Gaussian distribution law. Specifically, assuming that the scenario of extracting biometric features from one or more registration objects in the registration scenario satisfies the probability distribution model of the registration scenario, and currently it is necessary to infer the biometric registration of the first registration object in the registration scenario, it can be explained by formulas (1) to (3).

[0052] Formulas (1) to (3) are the following equations from top to bottom:

[0053] p(u)=N(u;0,εI)· (1)

[0054] p(x|u)=N(x;u,σI)· (2)

[0055] p(x)=∫p(x|u)p(u)du· (3)

[0056] In formulas (1) to (3), N represents a Gaussian distribution function; u represents the mean value of the biometric feature in the entire registration scenario; ε represents the probability distribution parameters between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; x represents the first registration object for biometric registration in the registration scenario; σ represents the probability distribution parameters between different samples of the same registration object included in the probability distribution model of the registration scenario.

[0057] Among them, the left side of formula (3) is the marginal probability, that is, it characterizes the distribution law of the biometric characteristics recorded after the first registered object performs biometric registration in the registration scenario. The right side of formula (3) is the expanded expression. It can be seen that formula (3) is a combination of formula (1) and formula (2) after expansion. Among them, formula (1) means to calculate through the Gaussian distribution function N, the mean u of the biometric characteristics in the entire registration scenario, and the probability distribution parameter ε between different registered objects. The left side of formula (2) is the posterior probability, which is the posterior probability of the first registered object x under the premise of calculating the mean u of the biometric characteristics in the entire registration scenario through formula (1). It should be understood that formulas (1) and (3) are both expressed as marginal probabilities, while formula (2) is expressed as posterior probabilities.

[0058] Please continue to refer to formulas (1) to (3), which are intended to illustrate that the scenario of extracting biometrics from one or more registered objects in a registration scenario satisfies the probability distribution model of the registration scenario, and currently it is necessary to infer the biometric registration of the first registered object in the registration scenario. If it is necessary to illustrate that the scenario of extracting biometrics from one or more recognized objects in a recognition scenario satisfies the probability distribution model of the recognition scenario, and currently it is necessary to infer the biometric registration of the first recognized object in the recognition scenario, it is only necessary to make corresponding parameter substitutions. Specifically, it can be used represents the mean value of the biometric features in each recognition scenario, represents a first identification object for biometric identification in the identification scenario, representing probability distribution parameters between different recognition objects included in the probability distribution model of the recognition scene, and The probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene are represented. That is, two sets of markers are used to distinguish the registration scene from the recognition scene, where the markers in formulas (1) to (3) correspond to the registration scene, and and Corresponding recognition scene.

[0059] In actual applications, when the first registrant performs biometric registration in a registration scenario, it is essentially a process of making predictions based on the posterior probability distribution template, which can be referred to in formula (4).

[0060] Formula (4) is the following equation:

[0061]

[0062] In formula (4), u k represents the biometrics recorded after the first registrant performs biometric registration in the registration scenario; N represents a Gaussian distribution function; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario. Formula (4) is the posterior probability distribution template, which means that after knowing multiple different registered objects until On the premise of the biometric feature, the first registration object records the biometric feature u after performing biometric registration in the registration scenario. k According to the posterior probability distribution template represented by formula (4), the corresponding posterior probability can be calculated by knowing the parameters in the above formula (4).

[0063] In the application, after the first registration object performs biometric registration in the registration scenario, it is known that the first registration object performing biometric registration in the registration scenario is x. Then, formulas (1) to (4) can be combined to calculate. When multiple different registration objects are known until On the premise of the biometric feature, the probability that the first registered object is x after the biometric feature registration in the registration scenario is, refer to formula (5).

[0064] Formula (5) is the following equation:

[0065]

[0066] In formula (5), N represents the Gaussian distribution function; u k represents the biometrics recorded after the first registrant performs biometric registration in the registration scenario; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected by the first registration object for biometric registration; ε represents the probability distribution parameters between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; x represents the first registration object for biometric registration under the registration scenario; σ represents the probability distribution parameters between different samples of the same registration object included in the probability distribution model of the registration scenario; represents the vector mean of all biometric features registered by the first registration object in the registration scenario. Formula (5) is a combination of Formula (4) and Formula (2). According to Formula (5), it can be calculated that the multiple different registration objects that have been known until On the premise of the biometric feature, the probability that the first registered object after biometric registration in the registration scenario is x.

[0067] In order to verify the identity of the first registered object, it is necessary to compare the biometrics extracted from the first identified object with the biometrics of multiple registered objects with known identities during biometric identification, and calculate the maximum likelihood score. The identity of the registered object corresponding to the maximum likelihood score is the identity of the first identified object. This process can be simplified by normalized likelihood scoring (NL), see formula (6).

[0068] Formula (6) is the following equation:

[0069]

[0070] In formula (6), NL represents the normalized likelihood scoring function; x represents the first registration object that performs biometric registration in the registration scenario; k represents the specific category to which the biometric registered by the first registration object belongs; H0 represents the case where the first registration object x is classified as the specific category k; H1 represents the case where the first registration object x is classified as a category other than the specific category k. Through the normalized likelihood scoring function NL, a unified threshold can be set to determine the identity of the first identification object to be identified, while also reducing the impact of the large difference in likelihood scoring results between different biometrics of the same object to be identified on the threshold. It can be seen from formula (6) that the normalized likelihood scoring function NL, after expansion, includes the known multiple different registration objects calculated by formula (5). until Under the premise of the biometric feature, the probability p of the first registered object being x after biometric registration in the registration scenario k (x), and the distribution law of the biometrics recorded by the first registrant after the biometrics registration in the registration scenario calculated by formula (3). Wherein, formula (5) calculates the posterior probability, and formula (3) calculates the marginal probability, so the normalized likelihood score of formula (6) can be understood as including the posterior probability part and the marginal probability part. Wherein, the posterior probability part of the likelihood score is calculated by formula (5), and the marginal probability part of the likelihood score is calculated by formula (3).

[0071] The normalized likelihood score NL of formula (6) can be converted into formulas (7) and (8) based on the likelihood ratio calculation under Gaussian distribution.

[0072] Formula (7) is the following equation:

[0073]

[0074] Formula (8) is the following equation:

[0075]

[0076] In formulas (7) and (8), NL represents the normalized likelihood scoring function; It is a variable set for simplicity of expression; n k represents the total number of different samples collected by the first registrant for biometric registration; ε represents the probability distribution parameter between different registrants included in the probability distribution model of the registration scenario; x represents the first registrant for biometric registration under the registration scenario; σ represents the probability distribution parameter between different samples of the same registrant included in the probability distribution model of the registration scenario; represents the vector mean of all biometric features registered by the first registration object in the registration scenario. For the sake of simplicity, some expressions in formula (7) are integrated into variables Then, the likelihood scoring process of formula (6) can be simplified by formula (7) and (8), which is convenient for biometric comparison and identity verification.

[0077] In practical applications, in order to compensate for the scoring performance loss caused by scene mismatch, corrections can be made in the following manner and based on the mapping relationship. Because the mapping relationship is reciprocal, the likelihood scoring calculation process can be mapped from the registration scene to the recognition scene, or the likelihood scoring calculation process can be mapped from the recognition scene to the registration scene. The following first describes an embodiment of mapping the likelihood scoring calculation process from the registration scene to the recognition scene.

[0078] In this embodiment, step S104 includes step S1041: calculating the posterior probability distribution template on the probability distribution model of the registration scene according to the biometrics of the first registration object, mapping the posterior probability distribution template from the registration scene to the recognition scene, and obtaining the mapped posterior probability distribution template.

[0079] There is a mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. This mapping relationship can be expressed as a linear mapping function, refer to formula (9).

[0080] Formula (9) is the following equation:

[0081]

[0082] In formula (9), M and b are constants of the linear mapping function. The configuration of the optimized linear mapping function can be obtained through machine learning training to maximize the above mapping relationship. Formula (9) means the mapping from the registration scene to the recognition scene, where x represents the first registration object for biometric registration in the registration scene; It means mapping x to the biometric features of the recognition scene according to the above linear mapping function.

[0083] First, the posterior probability distribution template is calculated on the probability distribution model of the registration scene according to the biometrics of the first registration object. Then, according to the mapping relationship, the posterior probability distribution template is mapped from the registration scene to the recognition scene to obtain the mapped posterior probability distribution template. Here, the posterior probability distribution template is in the sense of the registration scene, that is, it refers to the above formula (4). Formula (4) is the posterior probability distribution template, which means that after knowing multiple different registration objects until On the premise of the biometric feature, the first registration object records the biometric feature u after performing biometric registration in the registration scenario. k The distribution law of .

[0084] In order to map the posterior probability distribution template in the registration scenario represented by formula (4) to the posterior probability distribution template in the recognition scenario to obtain the mapped posterior probability distribution template, it is necessary to combine formula (4) and formula (9) to obtain formula (10).

[0085] Formula (10) is the following equation:

[0086]

[0087] In formula (10), represents the mean value of the biometric features for biometric recognition in the entire recognition scenario; N represents the Gaussian distribution function; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario; M and b are constants of the linear mapping function. It should be understood that in formula (10), the constant M of the mapping function is expressed in the form of a matrix, and M T represents the transpose of the matrix. Formula (10) means the posterior probability distribution template after mapping. It can be seen that formula (10) corresponds to formula (4) but represents the posterior probability distribution law in the recognition scenario.

[0088] In this embodiment, step S104 further includes step S1042: calculating the posterior probability part on the probability distribution model of the recognition scene according to the mapped posterior probability distribution template.

[0089] In the sense of the registration scenario, starting from the posterior probability distribution template represented by formula (4), combined with formula (2), the posterior probability that the first registered object after biometric registration in the registration scenario is x can be obtained as represented by formula (5). To this end, the posterior probability part can be calculated on the probability distribution model of the recognition scenario based on the mapped posterior probability distribution template represented by formula (10). Specifically, the part of the expanded expression of formula (5) corresponding to formula (4) can be replaced by formula (10), thereby obtaining formula (11).

[0090] Formula (11) is the following equation:

[0091]

[0092] In formula (11), represents a first identification object for biometric identification in the identification scenario; represents the mean value of the biometric features for biometric recognition in the entire recognition scenario; N represents the Gaussian distribution function; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario; M and b are constants of the linear mapping function.

[0093] Thus, similar to the deduction of formula (5) from formula (4), formula (11) can be deduced from formula (10). Formula (11) means that the multiple different registered objects that have been known until On the premise of the biometric features, the first identification object after biometric identification in the identification scenario is probability.

[0094] In this embodiment, step S104 also includes step S1043: calculating the marginal probability part on the probability distribution model of the recognition scene according to the characteristics of the first recognition object, and obtaining the likelihood score according to the posterior probability part and the marginal probability part.

[0095] The calculation of the normalized likelihood score in the sense of the recognition scenario can refer to formula (6). The posterior probability part calculated by formula (11) based on the probability distribution model of the recognition scenario after the mapping of the posterior probability distribution template needs to be used to replace the posterior probability part calculated in the sense of the registration scenario in formula (6). Then, the distribution law of the biometric features recorded after the first recognition object is identified in the recognition scenario is calculated according to formula (3). The parameters and identifiers of the corresponding recognition scenario need to be used to replace the parameters and identifiers of the corresponding registration scenario in formula (3). The likelihood score obtained in this way is similar to the form of formula (6), except that the posterior probability part is calculated according to formula (11), and the marginal probability part is calculated by replacing the parameters and representations of the corresponding registration scenario in formula (3) with the parameters and identifiers of the corresponding recognition scenario.

[0096] The likelihood score NL in the sense of the recognition scene can be calculated based on the likelihood ratio under the Gaussian distribution and converted into formulas (12) and (13).

[0097] Formula (12) is the following equation:

[0098]

[0099] Formula (13) is the following equation:

[0100]

[0101] In formulas (12) and (13), represents a first identification object for biometric identification in the identification scenario; Representing probability distribution parameters between different recognition objects included in the probability distribution model of the recognition scene; The probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario; M and b are constants of the linear mapping function. For simplicity, some expressions in formula (12) are integrated into variables The formula (13) is used to illustrate. In this way, the calculation result of formula (12) can be used to represent the simplified likelihood score.

[0102] Step S106: judging whether the identity of the first identified object is consistent with the identity of the first registered object according to the likelihood score.

[0103] Among them, the likelihood score can be calculated by the simplified likelihood ratio shown in formula (12), or the above method can be used to refer to formula (6) to calculate the normalized likelihood score in the sense of the recognition scene. It should be understood that, in combination with formula (9) to formula (13), the calculation process of the likelihood score is mapped from the registration scene to the recognition scene, and the impact of the mismatch between the scenes is taken into account, which is conducive to improving the scoring performance. Specifically, it can be briefly summarized as: according to the biometric characteristics of the first registered object, the posterior probability distribution template is calculated on the probability distribution model of the registration scene; according to the mapping relationship, the posterior probability distribution template is mapped from the registration scene to the recognition scene to obtain the mapped posterior probability distribution template; according to the mapped posterior probability distribution template, the posterior probability part is calculated on the probability distribution model of the recognition scene; according to the characteristics of the first recognized object, the marginal probability part is calculated on the probability distribution model of the recognition scene; according to the posterior probability part and the marginal probability part, the likelihood score is obtained.

[0104] See also Figure 2 , Figure 2 FIG. 2 is a flow chart showing another embodiment of an identity authentication method provided by an embodiment of the present application. Figure 2 As shown, the method includes the following steps.

[0105] In practical applications, in order to compensate for the scoring performance loss caused by scene mismatch, corrections can be made in the following manner and according to the mapping relationship. Because the mapping relationship is reciprocal, the likelihood scoring calculation process can be mapped from the registration scene to the recognition scene, or the likelihood scoring calculation process can be mapped from the recognition scene to the registration scene. The following describes an embodiment of mapping the likelihood scoring calculation process from the recognition scene to the registration scene.

[0106] Step S202: obtaining a biometric feature extracted from a first registration object in a registration scenario and a biometric feature extracted from a first identification object in an identification scenario.

[0107] It should be understood that the biometric feature can be a voiceprint feature, a facial feature, a fingerprint feature, or other biometric features used for identity authentication and identification. The definition of the first registration object and the first identification object is a person who provides a biometric feature and whose respective identities correspond to the provided biometric feature. For example, taking the biometric feature as a voiceprint feature as an example, the first registration object is the user who registers the voiceprint feature in the registration scenario, or can be called the speaker in the registration scenario. In contrast, the first identification object is the user who performs voiceprint feature identification in the identification scenario, or can be called the speaker in the identification scenario.

[0108] Step S204: Calculate a likelihood score based on the biometrics of the first registration object and the biometrics of the first identification object, wherein the likelihood score includes a posterior probability part and a marginal probability part.

[0109] Among them, the posterior probability part is determined according to the biometric feature of the first registration object, the probability distribution model of the registration scene and the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. The marginal probability part is determined according to the biometric feature of the first recognition object, the probability distribution model of the recognition scene and the mapping relationship. The probability distribution model of the registration scene includes the probability distribution parameters between different registration objects and the probability distribution parameters between different samples of the same registration object when extracting biometric features from one or more registration objects in the registration scene. The probability distribution model of the recognition scene includes the probability distribution parameters between different recognition objects and the probability distribution parameters between different samples of the same recognition object when extracting biometric features from one or more recognition objects in the recognition scene. The probability distribution model of the registration scene can be understood as the distribution law of the biometric features extracted in the limited space of the registration scene. Taking the biometric feature as the voiceprint feature as an example, the probability distribution model of the registration scene can be understood as the distribution of the voiceprint features of the speaker extracted in the limited space of the registration scene, or can be understood as a model of the feature distribution space used to describe the voiceprint features of the entire registration scene. In addition, there is a mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. This mapping relationship can be reflected by a mapping function, which can be linear or nonlinear. A simple linear mapping function is taken as an example below. This simple linear mapping function can be expressed as a linear function with two constants.

[0110] In this embodiment, step S204 includes step S2041: mapping the biometric feature of the first identification object from the identification scene to the registration scene to obtain the mapped biometric feature of the first identification object.

[0111] There is a mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. This mapping relationship can be expressed as a linear mapping function, refer to formula (14).

[0112] Formula (14) is the following equation:

[0113]

[0114] In formula (14), M and b are constants of the linear mapping function. The optimal configuration of the linear mapping function can be obtained through machine learning training to maximize the above mapping relationship. Formula (14) means the mapping from the recognition scene to the registration scene, where represents the first identification object for biometric identification in the identification scenario; x represents The biometric features are mapped to the registration scene according to the above linear mapping function.

[0115] In this embodiment, step S204 also includes step S2042: calculating the posterior probability distribution template on the probability distribution model of the registration scene based on the biometric characteristics of the first registration object, and calculating the posterior probability part on the probability distribution model of the registration scene based on the posterior probability distribution template and the mapped biometric characteristics of the first identified object.

[0116] The posterior probability distribution template refers to formula (4) to obtain formula (15).

[0117] Formula (15) is the following equation:

[0118]

[0119] In formula (15), u k represents the biometrics recorded after the first registrant performs biometric registration in the registration scenario; N represents a Gaussian distribution function; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; Represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario. It should be understood that because the multiple different registered objects that have been known are obtained in the sense of the registration scenario, formula (15) and formula (4) are consistent in mathematical terms, but the role of formula (15) here should be understood as calculating the posterior probability part based on the posterior probability distribution template and the mapped biometric features of the first identified object on the probability distribution model of the registration scenario. Specifically, substitute the biometric features mapped to the registration scenario according to the linear mapping function calculated by formula (14) into formula (5), that is, use Substituting x in formula (5) and expanding formula (5) yields formula (16).

[0120] Formula (16) is the following equation:

[0121]

[0122] In formula (16), N represents the Gaussian distribution function; u k represents the biometrics recorded after the first registrant performs biometric registration in the registration scenario; until Indicates the biometric features of multiple different registered objects that have been known; n k represents the total number of different samples collected by the first registration object for biometric registration; ε represents the probability distribution parameters between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; x represents the first registration object for biometric registration under the registration scenario; σ represents the probability distribution parameters between different samples of the same registration object included in the probability distribution model of the registration scenario; A vector mean representing all biometric features registered by the first registration object in the registration scenario; represents the first identification object for biometric identification in the identification scenario; M and b are constants of the linear mapping function. It can be seen that formula (16) is based on formula (15) and formula (3) after expansion. Formula (16) means that the multiple different registration objects that have been known until On the premise of the biometric features, the first identification object after biometric identification in the identification scenario is probability.

[0123] In this embodiment, step S204 also includes step S2043: calculating the marginal probability part on the probability distribution model of the recognition scene according to the biological characteristics of the first recognition object, and obtaining the likelihood score according to the posterior probability part and the marginal probability part.

[0124] Refer to formula (6), use Substituting x in formula (6) and expanding formula (6), we can get the marginal probability part of the likelihood score by referring to formula (17).

[0125] Formula (17) is the following equation:

[0126]

[0127] In formula (17), represents a first identification object for biometric identification in the identification scenario; Representing probability distribution parameters between different recognition objects included in the probability distribution model of the recognition scene; The probability distribution model represents the probability distribution parameters between different samples of the same recognition object included in the recognition scene.

[0128] The normalized likelihood score can refer to formula (6), except that the posterior probability part is calculated according to formula (16), and the marginal probability part is calculated according to formula (17).

[0129] The likelihood score NL in the sense of the registration scenario can be calculated based on the likelihood ratio under the Gaussian distribution and converted into formula (18).

[0130] Formula (18) is the following equation:

[0131]

[0132] In formula (18), represents a first identification object for biometric identification in the identification scenario; k represents a specific category to which the biometric identified by the first identification object belongs; Representing probability distribution parameters between different recognition objects included in the probability distribution model of the recognition scene; The probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario; M and b are constants of the linear mapping function.

[0133] Step S206: judging whether the identity of the first identified object is consistent with the identity of the first registered object according to the likelihood score.

[0134] Among them, the likelihood score can be calculated by the simplified likelihood ratio shown in formula (18), or the above method can be used to refer to formula (6) to calculate the normalized likelihood score NL in the sense of the registration scene. It should be understood that, in combination with formula (14) to formula (18), the calculation process of the likelihood score is mapped from the recognition scene to the registration scene, and the impact of the mismatch between the scenes is taken into account, which is conducive to improving the scoring performance. Specifically, it can be briefly summarized as: according to the mapping relationship, the biometric characteristics of the first recognition object are mapped from the recognition scene to the registration scene to obtain the mapped biometric characteristics of the first recognition object; the posterior probability distribution template is calculated based on the biometric characteristics of the first registration object on the probability distribution model of the registration scene; the posterior probability part is calculated based on the posterior probability distribution template and the mapped biometric characteristics of the first recognition object on the probability distribution model of the registration scene; the marginal probability part is calculated based on the biometric characteristics of the first recognition object on the probability distribution model of the recognition scene; the likelihood score is obtained based on the posterior probability part and the marginal probability part.

[0135] In a possible implementation, the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene is as shown in formula (14). And when the probability distribution parameters ε between different registration objects included in the probability distribution model of the registration scenario are all small, the linear mapping relationship of formula (14) can be simplified to a global translation, refer to formula (19).

[0136] Formula (19) is the following equation:

[0137]

[0138] In formula (19), b is the global displacement. The configuration of the optimal linear mapping function can be obtained through machine learning training to maximize the above mapping relationship. Formula (19) means the mapping from the recognition scene to the registration scene, where represents the first identification object for biometric identification in the identification scenario; x represents The biometric features are mapped to the registration scene according to the above linear mapping function.

[0139] According to formula (19), combined with formulas (15) to (18), a simplified likelihood ratio calculation can be obtained, refer to formula (20).

[0140] Formula (20) is the following equation:

[0141]

[0142] In formula (20), represents a first identification object for biometric identification in the identification scenario; k represents a specific category to which the biometric identified by the first identification object belongs; Representing probability distribution parameters between different recognition objects included in the probability distribution model of the recognition scene; The probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene; n k represents the total number of different samples collected for biometric registration of the first registration object; ε represents the probability distribution parameter between different registration objects included in the probability distribution model of the registration scenario; I represents a conventional diagonal matrix; represents the vector mean of all biometric features registered by the first registered object in the registration scenario; σ represents the probability distribution parameter between different samples of the same registered object included in the probability distribution model of the registration scenario; b is the global displacement.

[0143] Please continue reading Figure 1 and Figure 2 , the mapping function used to reflect the inverse mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene can be a linear analytical solution, or a nonlinear mapping relationship obtained by a numerical solution or a neural network algorithm. The following is a brief description of the method for training the mapping function: estimate the feature distribution in the registration scene according to the registration training data; estimate the feature distribution in the recognition scene according to the recognition training data; calculate the loss function based on the maximum likelihood criterion estimation according to the mapping function, the registration training data and the recognition training data; wherein the mapping function is used for the inverse mapping between the feature distribution in the registration scene and the feature distribution in the recognition scene; adjust the parameters of the mapping function according to the loss function until the loss function meets the optimization condition of minimization, and obtain the trained mapping function.

[0144] See also Figure 2and formula (14), where M and b are constants of the linear mapping function, the configuration of the optimized linear mapping function can be obtained through machine learning training to maximize the above mapping relationship. To this end, for the mapping function of formula (14), a loss function shown in formula (21) can be designed.

[0145] Formula (21) is the following equation:

[0146]

[0147] In formula (21), represents the first identification object for biometric identification in the identification scenario; K represents the total number of different registered objects; N represents the total number of samples of the biometrics of each registered object. Through the loss function shown in formula (21), the purpose of learning the optimal mapping function can be achieved through a machine learning model for the maximum likelihood linear regression (MLLR) task or a training method based on the maximum likelihood criterion estimation.

[0148] Please continue reading Figure 1 and Figure 2 In a possible implementation, it is necessary to consider the changes in the biometric distribution caused by the changes in the identification object itself, and it is necessary to update the relevant model according to these changes, so as to correct the likelihood probability score of the mismatch between the registration and identification scenes caused by the changes in the identification object itself. For example, in the process of biometric identification, high-probability biometrics that have passed can be screened, and the biometric distribution of a specific registration object can be updated using a dynamic update method. The specific steps include: obtaining biometrics extracted from multiple samples of the second identification object in the identification scene and calculating the likelihood scores corresponding to each of the multiple samples of the second identification object; based on the likelihood scores corresponding to each of the multiple samples of the second identification object, screening out features for state update from the biometrics extracted from the multiple samples of the second identification object; using the features for state update, updating the probability distribution parameters between different samples of the same identification object included in the probability distribution model of the identification scene. The revised likelihood score refers to formula (22).

[0149] Formula (22) is the following equation:

[0150]

[0151] In formula (22), NL represents the normalized likelihood scoring function; is a variable set for simplicity of expression, refer to formula (8); n krepresents the total number of different samples collected by the first registrant for biometric registration; ε represents the probability distribution parameter between different registrants included in the probability distribution model of the registration scenario; x represents the first registrant for biometric registration under the registration scenario; σ represents the probability distribution parameter between different samples of the same registrant included in the probability distribution model of the registration scenario; The probability distribution parameters between different samples of the same identification object included in the probability distribution model of the identification scene can be represented. The modified likelihood score can be obtained through formula (22), thereby adjusting and updating the distribution parameters according to the biometric features of multiple samples of the second identification object, which is conducive to dynamic adjustment and update according to the changes of the identification object itself.

[0152] In some exemplary embodiments, in order to reflect the likelihood scoring results corresponding to different states of the same object, a clustering algorithm can be used in the process of continuously collecting the user's biometric features, clustering the user's biometric features in the updated feature pool, and maintaining multiple models and distributions for the user. Common clustering algorithms include k-means algorithm (k-means), Gaussian Mixture Model (Gaussian Mixture Model, GMM), etc. It can be understood that the user sub-models clustered according to the clustering algorithm can represent different states of a specific user speaking. When identifying, the results of the scoring of multiple sub-models are fused to obtain a more stable recognition result. The specific steps include: clustering the probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene through a clustering algorithm and obtaining multiple sub-models; the multiple sub-models each correspond to a different state of the same recognition object; the marginal probability part of the likelihood scoring fuses the scoring results of each of the multiple sub-models. In this way, the scoring fusion is achieved through the clustering algorithm, which is conducive to making adjustments according to the different states of the same recognition object.

[0153] Figure 3 A schematic diagram of identity authentication based on registration and identification of biometric features provided in an embodiment of the present application is shown. It should be understood that the biometric feature can be a voiceprint feature, a facial feature, a fingerprint feature, or other biometric features used for identity authentication and identification. The definition of the registration object and the identification object is a person who provides a biometric feature and whose respective identity corresponds to the provided biometric feature. For example, taking the biometric feature as a voiceprint feature as an example, the registration object is the user who registers the voiceprint feature in the registration scenario, or can be called the speaker in the registration scenario.

[0154] Figure 3The distribution difference mapping 350 in refers to the distribution difference between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. This distribution difference leads to the need to make corrections in the scoring decision 360. To this end, it is necessary to map the calculation process of the likelihood score from the registration scene to the recognition scene, or to map the calculation process of the likelihood score from the recognition scene to the registration scene through the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. In other words, it is necessary to adjust the scoring decision 360 according to the distribution difference mapping 350. Assuming that this mapping relationship can be expressed as a linear mapping function, refer to formula (9) or (14). Therefore, the calculation process of the likelihood score can be mapped from the registration scene to the recognition scene in combination with formula (9) to formula (13), or the calculation process of the likelihood score can be mapped from the recognition scene to the registration scene in combination with formula (14) to formula (18).

[0155] Please continue reading Figure 3 , the registration scene and the recognition scene can be understood as two relatively speaking environments, each corresponding to a different feature space distribution. In the registration stage, multiple registered objects register biometrics 310, and then the registered biometrics are extracted 312. Based on the extracted biometrics, a biometric model 314 of the registered object is established. The biometric model can also be understood as a user profile, which is used to correspond the user's identity to his biometrics one by one, so as to verify the identity and identify the biometrics during the recognition stage. In the recognition stage, the object to be recognized recognizes the biometrics 320, extracts the biometrics from the recognized biometrics of the object to be recognized, and then performs a scoring judgment 360. When the registration scene and the recognition scene do not match, that is, mismatch, this means that the probability distribution model or probability distribution function of the feature space distribution of the biometrics in the registration scene is different from the probability distribution model or distribution function of the feature space distribution corresponding to the recognition scene. For example, assuming that the biometrics of an object in the registration scene and the recognition scene meet the Gaussian distribution law and can be expressed as a specific Gaussian distribution function, then the mean and variance parameters of the Gaussian distribution function are different from other Gaussian distribution functions. In addition, the biometric features of different objects in the same feature space can also be expressed as a specific Gaussian distribution function with corresponding mean and variance parameters. The mismatch between the registration scene and the recognition scene can also be understood as the statistical inconsistency of the feature spaces of the registration scene and the recognition scene, which makes the direct similarity comparison ineffective and easily causes performance loss. The reasons for the mismatch between the registration scene and the recognition scene are diverse. They may be caused by external influences, such as channels, environments, sound pickup distance, etc., or they may be caused by changes in the user or speaker himself, such as changes in the speaker's voice over time, changes in the way of speaking, etc.

[0156] Please continue reading Figure 3In order to cope with the impact of the mismatch between the above-mentioned registration scene and the recognition scene, the registration scene development data 330 is used to estimate the registration scene distribution parameter 332, that is, the distribution parameter of the biometric features in the feature space corresponding to the registration scene, and the recognition scene development data 340 is also used to estimate the recognition scene distribution parameter 342, that is, the distribution parameter of the biometric features in the feature space corresponding to the recognition scene. Specifically, the development data adapted to the registration scene and the development data adapted to the recognition scene are used to obtain the initial probability distribution model of the registration scene and the initial probability distribution model of the recognition scene respectively; according to the initial probability distribution model of the registration scene and the initial probability distribution model of the recognition scene, a distribution difference mapping 350 is obtained, for example, expressed as a corresponding distribution difference mapping function. Then, the distribution difference mapping function is trained using the maximum likelihood criterion estimation to obtain a trained distribution difference mapping function; the trained distribution difference mapping function is used to improve the scoring decision 360. That is to say, firstly, the distribution of biometric features in the registration scene and the recognition scene is modeled using the development data of multiple scenes; then, the distribution difference mapping model of the mismatch between the registration scene and the recognition scene is learned using the machine learning method; when identifying the user, the user features of the registration stage and the recognition stage are mapped to the same matching spatial distribution using the established distribution difference mapping model for similarity comparison, thereby solving the impact of the mismatch between the registration scene and the recognition scene. On the other hand, the user biometric model 314 established in the registration stage can be mapped to the space matching the recognition scene in combination with the function corresponding to the distribution difference mapping 350 to obtain the mapped biometric model 316, and then a scoring decision 360 is performed. In a possible implementation, the function corresponding to the distribution difference mapping 350 can be designed using the above formula (22).

[0157] Figure 4 The structural block diagram of the identity authentication system provided by the embodiment of the present application is shown. It should be understood that the biometric feature can be a voiceprint feature, a facial feature, a fingerprint feature, or other biometric features used for identity authentication and identification. The definition of the registration object and the identification object is a person who provides a biometric feature and whose respective identity corresponds to the provided biometric feature. For example, taking the biometric feature as a voiceprint feature as an example, the registration object is the user who registers the voiceprint feature in the registration scenario, or can be called the speaker in the registration scenario.

[0158] Figure 4 The distribution difference training module 444 in the example refers to a mapping relationship that reflects the statistical distribution difference of the feature spaces of the registration scene and the recognition scene obtained through deep machine learning technology, such as a neural network algorithm. Figure 4As shown, the identity authentication system 400 includes a registration module 410, an identification module 420, a biometric model storage device 430, a multi-device development data storage device 440, a biometric distribution estimation module 442, and a distribution difference training module 444. Among them, the registration module 410 can be deployed on the registration device to allow users to register biometrics or can be deployed on the cloud to allow users to register, including the following steps: collecting user biometrics through the registration device, extracting biometrics from them through a feature extractor, building a model based on the extracted biometrics and binding the user ID or other identity credentials, and finally transmitting the biometric model obtained by the user through the registration device to the biometric model storage device 430 for network cloud storage. Among them, the extractor can be a deep neural network model trained with a large number of user voices unrelated to the registration device. The identification module 420 can be deployed on the identification device to allow users to perform biometric identification, including the following steps: collecting user biometrics through the identification device, extracting biometrics from them through the extractor, loading the user registered biometric model from the connected cloud storage, and finally performing likelihood scoring judgment. Among them, the identification module 420 can also make a judgment on the matching of the user identification device and the registration device in advance, that is, to judge the factors that may affect the scoring performance, such as the channel difference between the two devices. When it is judged that the registration device and the identification device match, a judgment can be made directly based on the likelihood score; and when it is judged that the registration device and the identification device do not match, the pre-trained distribution difference mapping model can be called to perform a modified likelihood score calculation, and finally a judgment is made based on the modified likelihood score. Among them, the modified likelihood score calculation can be performed by the scoring judgment module 422.

[0159] Please continue reading Figure 4 , the biometric model storage device 430 includes a biometric model storage device 432, a multi-device biometric distribution storage device 434 and a multi-device distribution mapping storage device 436. Among them, the biometric model storage device 432 is communicatively connected to the registration module 410 and receives the biometric model obtained by the user through the registration device registration. The multi-device biometric distribution storage device 434 and the multi-device distribution mapping storage device 436 are connected to the distribution difference training module 444, and provide pre-trained biometric distribution and distribution mapping models for multiple devices. The distribution difference training module 444 trains the distribution difference mapping function with the registration device for the device that needs to perform cross-device biometric recognition, for example, through maximum likelihood criterion estimation. Specifically, the multi-scenario development data stored in the multi-device development data storage device 440 can be loaded, and after the distribution of biometrics in the registration scenario and the recognition scenario is modeled by the feature distribution estimation module 442, the distribution difference training module 444 is used to learn the mismatched distribution difference mapping model in the registration scenario and the recognition scenario, thereby solving the impact of the mismatch between the registration scenario and the recognition scenario. It should be understood that, Figure 4The multi-device biometric distribution memory 434 and the multi-device distribution mapping memory 436 are shown for cross-device, i.e., multi-device or multi-channel application scenarios. The pre-trained biometric distribution and distribution mapping models for multiple devices can also take into account the mismatch in such cross-device or multi-channel scenarios.

[0160] Please continue reading Figure 4 , the registration module 410, the identification module 420 and the biometric model storage device 430 can be deployed in different places, or can be located in the same place, or other possible combinations. For example, the registration module 410 and the identification module 420 can be located in the registration scene and the identification scene respectively, and the biometric model storage device 430 can be a cloud storage device or a local storage device set in the cloud. When the biometric model storage device 430 in the form of cloud storage is used to save the biometric distribution and distribution mapping function scenarios of multiple devices, it can be beneficial to separate the registration device and the identification device, and facilitate the completion of registration and identification at different places and times through the biometric model storage device 430, thereby improving system efficiency. On the other hand, the multi-device development data storage device 440, the biometric distribution estimation module 442 and the distribution difference training module 444 can be set separately from the biometric model storage device 430, so that the training process can be completed independently. In addition, it should be understood that, Figure 4 The scoring and judging module 422 is shown to be deployed in the identification module 420, because in most application scenarios, it is necessary to provide a likelihood scoring result and perform identity authentication after the identification module 420 collects the user's biometric features. However, in a possible implementation, the scoring and judging module 422 can also be set up separately, that is, the identification module 420 only collects the biometric features in the identification scenario, and the biometric features in the identification scenario collected centrally can be centrally processed by the scoring and judging module 422 at a certain time period thereafter.

[0161] Figure 4 The identity system 400 shown can be used to implement Figure 1 or Figure 2Specifically, the registration module 410 is used to obtain a biometric feature extracted from a first registration object in a registration scenario; the identification module 420 is used to obtain a biometric feature extracted from a first identification object in an identification scenario; the scoring judgment module 422 is used to calculate a likelihood score based on the biometric feature of the first registration object and the biometric feature of the first identification object and to judge whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score. Among them, the likelihood scoring includes a posterior probability part and a marginal probability part; the posterior probability part is determined according to the biometric characteristics of the first registered object, the probability distribution model of the registration scenario and the mapping relationship between the probability distribution model of the registration scenario and the probability distribution model of the recognition scenario; the probability distribution model of the registration scenario includes the probability distribution parameters between different registered objects and the probability distribution parameters between different samples of the same registered object when extracting biometric characteristics from one or more registered objects in the registration scenario, and the probability distribution model of the recognition scenario includes the probability distribution parameters between different recognized objects and the probability distribution parameters between different samples of the same recognized object when extracting biometric characteristics from one or more recognized objects in the recognition scenario; the marginal probability part is determined according to the biometric characteristics of the first recognized object, the probability distribution model of the recognition scenario and the mapping relationship.

[0162] Figure 5 The structural block diagram of the identity authentication system for dynamically matching the changes of the user himself provided by the embodiment of the present application is shown. It should be understood that the biometric feature can be a voiceprint feature, a facial feature, a fingerprint feature or other biometric features used for identity authentication and identification. The definition of the registration object and the identification object is a person who provides a biometric feature and whose respective identity corresponds to the provided biometric feature. For example, taking the biometric feature as a voiceprint feature as an example, the registration object is the user who registers the voiceprint feature in the registration scenario, or can be called the speaker in the registration scenario.

[0163] Figure 5 The scoring and judging module 522 can realize Figure 1 or Figure 2 The calculation process of likelihood scoring is shown in Figure 2. In actual applications, changes in the user may cause mismatches between the registration scene and the recognition scene. Considering that the user's state changes in real time and is affected by the user's emotions, posture, and speech speed, which causes changes in the biometric distribution model, it is necessary to dynamically match the biometric distribution model of the user's current state. Figure 5The scoring decision module 522 combined with the update decision module 540 can screen out the user features that need to be updated and store them in the to-be-updated feature memory 542. After the user features are updated by the biometric feature distribution re-estimation module 544, the updated user personal features are stored in the personal feature distribution memory 534, thereby realizing the correction of the likelihood scoring result provided by the recognition module 520.

[0164] Please continue reading Figure 5 , the identity authentication system 500 includes a registration module 510, an identification module 520, a biometric model storage device 530, a personal feature distribution storage device 534, an update judgment module 540, a feature storage device 542 to be updated, and a biometric distribution re-estimation module 544. Among them, the registration module 510 can be deployed on the registration device to allow users to register, including the following steps: collecting user biometrics through the registration device, extracting biometrics from them through a feature extractor, establishing a biometric model based on the extracted biometrics and binding the user ID or other identity credentials, and finally transmitting the biometric model obtained by the user through the registration device to the biometric model storage device 530 for storage. The identification module 520 can be deployed on the identification device to allow users to identify through voice or can be deployed in the cloud to allow users to identify, including the following steps: collecting user biometrics through the identification device, extracting biometrics from them through a feature extractor, downloading the biometric model registered by the user from the connected biometric model storage device 532, and finally performing likelihood scoring through a scoring judgment module. Considering that the user's biometrics is a user behavior feature and is easily affected by the user's speech content, time changes, and speaking methods. Moreover, only a few features are used when registering a user, so the registered biometric features can only represent the biometric features of the user at the time of registration. During recognition, changes in the user's own situation will also affect the recognition effect, which may lead to a mismatch between the registration scene and the recognition scene.

[0165] In some exemplary embodiments, the scoring judgment module 522 is also used to: obtain biometric features extracted from multiple samples of the second identification object in the identification scenario and calculate the likelihood scores corresponding to each of the multiple samples of the second identification object; screen out features for state updating from the biometric features extracted from the multiple samples of the second identification object based on the likelihood scores corresponding to each of the multiple samples of the second identification object; and use the features for state updating to update the probability distribution parameters between different samples of the same identification object included in the probability distribution model of the identification scenario.

[0166] The specific embodiments provided by the present application can be implemented with any one or combination of hardware, software, firmware or solid-state logic circuits, and can be implemented in combination with signal processing, control and / or special circuits. The equipment or device provided by the specific embodiments of the present application may include one or more processors (e.g., microprocessors, controllers, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), etc.), which process various computer-executable instructions to control the operation of the equipment or device. The equipment or device provided by the specific embodiments of the present application may include a system bus or a data transmission system that couples various components together. The system bus may include any one or a combination of different bus structures, such as a memory bus or a memory controller, a peripheral bus, a universal serial bus and / or a processor or local bus utilizing any one of a variety of bus architectures. The equipment or device provided by the specific embodiments of the present application may be provided separately, may be a part of a system, or may be a part of other equipment or devices.

[0167] The specific embodiments provided in this application may include a computer-readable storage medium or be combined with a computer-readable storage medium, such as one or more storage devices capable of providing non-temporary data storage. The computer-readable storage medium / storage device may be configured to store data, programmers and / or instructions, which, when executed by a processor of a device or apparatus provided in a specific embodiment of this application, enable these devices or apparatuses to perform related operations. The computer-readable storage medium / storage device may include one or more of the following features: volatility, non-volatility, dynamic, static, readable / writable, read-only, random access, sequential access, location addressability, file addressability, and content addressability. In one or more exemplary embodiments, the computer-readable storage medium / storage device may be integrated into the device or apparatus provided in a specific embodiment of this application or belong to a common system. Computer-readable storage media / storage devices may include optical storage devices, semiconductor storage devices and / or magnetic storage devices, etc., and may also include random access memory (RAM), flash memory, read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, recordable and / or rewritable compact disks (CDs), digital versatile disks (DVDs), mass storage media devices or any other form of suitable storage media.

[0168] The above is the implementation method of the embodiment of the present application. It should be noted that the steps in the method described in the specific embodiment of the present application can be adjusted in order, merged and deleted according to actual needs. In the above embodiments, the description of each embodiment has its own emphasis. For the part that is not described in detail in a certain embodiment, refer to the relevant description of other embodiments. It is understandable that the structure shown in the embodiment of the present application and the accompanying drawings does not constitute a specific limitation on the relevant device or system. In other embodiments of the present application, the relevant device or system may include more or fewer components than the specific embodiments and the accompanying drawings, or combine certain components, or split certain components, or have different component arrangements. Those skilled in the art will understand that, without departing from the spirit and scope of the specific embodiments of the present application, various modifications or changes can be made to the arrangement, operation and details of the methods and equipment recorded in the specific embodiments; without departing from the principles of the embodiments of the present application, several improvements and modifications can also be made, and these improvements and modifications are also regarded as the scope of protection of the present application.

Claims

1. A method of identity verification, It is characterized in that The method comprises: Obtaining a biometric feature extracted from a first registration object in a registration scenario; Obtaining a biometric feature extracted from a first identification object in an identification scenario; Calculating a likelihood score based on the biometrics of the first registration object and the biometrics of the first identification object; and determining whether the identity of the first identified object is consistent with the identity of the first registered object according to the likelihood score, The likelihood score includes a posterior probability part and a marginal probability part. The posterior probability part is determined according to the biometrics of the first registered object, the probability distribution model of the registration scene, and the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. The probability distribution model of the registration scenario includes probability distribution parameters between different registration objects and probability distribution parameters between different samples of the same registration object when extracting biometric features from one or more registration objects in the registration scenario, and the probability distribution model of the recognition scenario includes probability distribution parameters between different recognition objects and probability distribution parameters between different samples of the same recognition object when extracting biometric features from one or more recognition objects in the recognition scenario, The edge probability part is determined according to the biological characteristics of the first recognition object, the probability distribution model of the recognition scene and the mapping relationship.

2. The method according to claim 1, It is characterized in that Calculating the likelihood score according to the biometric feature of the first registration object and the biometric feature of the first identification object includes: According to the mapping relationship, the biometric feature of the first identification object is mapped from the identification scene to the registration scene to obtain the mapped biometric feature of the first identification object, Calculating a posterior probability distribution template on the probability distribution model of the registration scene according to the biometric feature of the first registration object, The posterior probability part is calculated on the probability distribution model of the registration scene according to the posterior probability distribution template and the mapped biometric feature of the first recognition object, Calculating the marginal probability part on the probability distribution model of the recognition scene according to the biological feature of the first recognition object, The likelihood score is obtained according to the posterior probability part and the marginal probability part.

3. The method according to claim 1, It is characterized in that Calculating the likelihood score according to the biometric feature of the first registration object and the biometric feature of the first identification object includes: Calculating a posterior probability distribution template on the probability distribution model of the registration scene according to the biometric feature of the first registration object, According to the mapping relationship, the posterior probability distribution template is mapped from the registration scene to the recognition scene to obtain a mapped posterior probability distribution template, Calculating the posterior probability part on the probability distribution model of the recognition scene according to the mapped posterior probability distribution template, Calculating the edge probability part on the probability distribution model of the recognition scene according to the characteristics of the first recognition object, The likelihood score is obtained according to the posterior probability part and the marginal probability part.

4. The method according to claim 2 or 3, It is characterized in that The probability distribution model of the registration scene and the probability distribution model of the recognition scene both obey linear Gaussian distribution, The probability distribution parameters between different registered objects include the mean and variance of the feature space distribution of different registered objects in the registration scene, The probability distribution parameters between different samples of the same registered object include the variance of the feature space distribution of the same registered object in the registration scene, The probability distribution parameters between different recognition objects include the mean and variance of the feature space distribution of different recognition objects in the recognition scene, The probability distribution parameters between different samples of the same recognition object include the variance of the feature space distribution of the same recognition object in the recognition scene.

5. The method according to claim 4, It is characterized in that The likelihood score is calculated based on the linear Gaussian distribution.

6. The method according to claim 2 or 3, It is characterized in that The mapping relationship is determined according to a trained mapping function, and the trained mapping function is obtained by the following method: Estimate the feature distribution in the registration scenario according to the registered training data, Estimate the feature distribution in the recognition scenario according to the recognition training data, A loss function based on maximum likelihood criterion estimation is calculated according to a mapping function, the registration training data and the recognition training data, wherein the mapping function is used for inverse mapping between the feature distribution in the registration scenario and the feature distribution in the recognition scenario, The parameters of the mapping function are adjusted according to the loss function until the loss function satisfies the optimization condition of minimization, thereby obtaining the trained mapping function.

7. The method according to claim 6, It is characterized in that The mapping function is a linear function of one variable, and the parameters of the mapping function include two constants of the linear function of one variable.

8. The method according to claim 6, It is characterized in that The machine learning model corresponding to the loss function is used for the maximum likelihood linear regression MLLR task.

9. The method according to any one of claims 1 to 3, It is characterized in that The method further comprises: Obtaining biometric features extracted from multiple samples of the second identification object in the identification scenario and calculating likelihood scores corresponding to each of the multiple samples of the second identification object, According to the likelihood scores corresponding to the plurality of samples of the second identification object, the features for state update are screened out from the biometric features extracted from the plurality of samples of the second identification object, The feature for state updating is used to update probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene.

10. The method according to any one of claims 1 to 3, It is characterized in that The method further comprises: By using a clustering algorithm, probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene are clustered to obtain multiple sub-models. Wherein, the multiple sub-models each correspond to a different state of the same recognition object; Among them, the marginal probability part of the likelihood scoring integrates the scoring results of each of the multiple sub-models.

11. The method according to any one of claims 1 to 3, It is characterized in that The biometric feature includes a voiceprint feature.

12. The method according to any one of claims 1 to 3, It is characterized in that The biometric features include facial features.

13. The method according to any one of claims 1 to 3, It is characterized in that The biometric feature includes a fingerprint feature.

14. The method according to any one of claims 1 to 3, It is characterized in that The registration scenario and the identification scenario correspond to different devices respectively.

15. A chip system, It is characterized in that The chip system is applied to an electronic device; the chip system includes one or more interface circuits and one or more processors; the interface circuit and the processor are interconnected through lines; the interface circuit is used to receive a signal from the memory of the electronic device and send the signal to the processor, the signal including computer instructions stored in the memory; when the processor executes the computer instructions, the electronic device executes the method as described in any one of claims 1 to 14.

16. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer program instructions, which, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 14.

17. An identity verification system, It is characterized in that The identity verification system comprises: A registration module, configured to obtain a biometric feature extracted from a first registration object in a registration scenario; A recognition module, used to obtain a biometric feature extracted from a first recognition object in a recognition scenario; as well as a scoring judgment module, configured to calculate a likelihood score based on the biometric features of the first registration object and the biometric features of the first identification object, and to judge whether the identity of the first identification object is consistent with the identity of the first registration object based on the likelihood score, The likelihood score includes a posterior probability part and a marginal probability part. The posterior probability part is determined according to the biometrics of the first registered object, the probability distribution model of the registration scene, and the mapping relationship between the probability distribution model of the registration scene and the probability distribution model of the recognition scene. The probability distribution model of the registration scenario includes probability distribution parameters between different registration objects and probability distribution parameters between different samples of the same registration object when extracting biometric features from one or more registration objects in the registration scenario, and the probability distribution model of the recognition scenario includes probability distribution parameters between different recognition objects and probability distribution parameters between different samples of the same recognition object when extracting biometric features from one or more recognition objects in the recognition scenario, The edge probability part is determined according to the biological characteristics of the first recognition object, the probability distribution model of the recognition scene and the mapping relationship.

18. The identity verification system according to claim 17, It is characterized in that The scoring and judging module is specifically used for: According to the mapping relationship, the biometric feature of the first identification object is mapped from the identification scene to the registration scene to obtain the mapped biometric feature of the first identification object, Calculating a posterior probability distribution template on the probability distribution model of the registration scene according to the biometric feature of the first registration object, The posterior probability part is calculated on the probability distribution model of the registration scene according to the posterior probability distribution template and the mapped biometric feature of the first recognition object, Calculating the marginal probability part on the probability distribution model of the recognition scene according to the biological feature of the first recognition object, The likelihood score is obtained according to the posterior probability part and the marginal probability part.

19. The identity verification system according to claim 17, It is characterized in that The scoring and judging module is specifically used for: Calculating a posterior probability distribution template on the probability distribution model of the registration scene according to the biometric feature of the first registration object, According to the mapping relationship, the posterior probability distribution template is mapped from the registration scene to the recognition scene to obtain a mapped posterior probability distribution template, Calculating the posterior probability part on the probability distribution model of the recognition scene according to the mapped posterior probability distribution template, Calculating the edge probability part on the probability distribution model of the recognition scene according to the characteristics of the first recognition object, The likelihood score is obtained according to the posterior probability part and the marginal probability part.

20. The identity verification system according to claim 18 or 19, It is characterized in that The probability distribution model of the registration scene and the probability distribution model of the recognition scene both obey linear Gaussian distribution, The probability distribution parameters between different registered objects include the mean and variance of the feature space distribution of different registered objects in the registration scene, The probability distribution parameters between different samples of the same registered object include the variance of the feature space distribution of the same registered object in the registration scene, The probability distribution parameters between different recognition objects include the mean and variance of the feature space distribution of different recognition objects in the recognition scene, The probability distribution parameters between different samples of the same recognition object include the variance of the feature space distribution of the same recognition object in the recognition scene.

21. The identity verification system according to claim 20, It is characterized in that The likelihood score is calculated based on the linear Gaussian distribution.

22. An identity verification system according to claim 18 or 19, It is characterized in that The mapping relationship is determined according to a trained mapping function, and the trained mapping function is obtained by the following method: Estimate the feature distribution in the registration scenario according to the registered training data, Estimate the feature distribution in the recognition scenario according to the recognition training data, A loss function based on maximum likelihood criterion estimation is calculated according to a mapping function, the registration training data and the recognition training data, wherein the mapping function is used for inverse mapping between the feature distribution in the registration scenario and the feature distribution in the recognition scenario, The parameters of the mapping function are adjusted according to the loss function until the loss function satisfies the optimization condition of minimization, thereby obtaining the trained mapping function.

23. The identity verification system according to claim 22, It is characterized in that The mapping function is a linear function of one variable, and the parameters of the mapping function include two constants of the linear function of one variable.

24. The identity verification system according to claim 22, It is characterized in that The machine learning model corresponding to the loss function is used for the maximum likelihood linear regression MLLR task.

25. An identity verification system according to any one of claims 17 to 19, It is characterized in that The scoring and judging module is also used for: Obtaining biometric features extracted from multiple samples of the second identification object in the identification scenario and calculating likelihood scores corresponding to each of the multiple samples of the second identification object, According to the likelihood scores corresponding to the plurality of samples of the second identification object, the features for state update are screened out from the biometric features extracted from the plurality of samples of the second identification object, The feature for state updating is used to update probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene.

26. An identity verification system according to any one of claims 17 to 19, It is characterized in that The scoring and judging module is also used for: By using a clustering algorithm, probability distribution parameters between different samples of the same recognition object included in the probability distribution model of the recognition scene are clustered to obtain multiple sub-models. Wherein, the multiple sub-models each correspond to a different state of the same recognition object; Among them, the marginal probability part of the likelihood scoring integrates the scoring results of each of the multiple sub-models.

Citation Information

Cited By

  • Identity verification method and system

    WO2022110782A1