Method for updating biometric feature recognition model and related product

Abstract

The invention discloses a method for updating a biological feature recognition model and a related product. The method comprises the following steps: in response to a demand of updating an old biological feature recognition model into a new biological feature recognition model, performing style migration processing on a registered biological feature corresponding to the old biological feature recognition model to obtain a new learning feature set related to a biological feature style corresponding to the new biological feature recognition model; processing the to-be-recognized image by using the new biological feature recognition model in response to an obtained biological feature recognition requirement for the to-be-recognized image to obtain a new model feature; and selectively replacing the new learning features in the new learning feature set with the new model features so as to realize registration feature updating of the new biological feature recognition model. Through the technical scheme disclosed by the invention, the non-inductive upgrading of the model can be realized, and meanwhile, the precision loss in the model updating process can be effectively reduced, so that the identification accuracy of the model is improved.

Description

Technical Field

[0001] This disclosure generally relates to the field of biometric recognition technology. More specifically, this disclosure relates to a method for updating a biometric recognition model, as well as an electronic device and a computer-readable storage medium for performing the foregoing method. Background Art

[0002] Biometric recognition algorithms such as face recognition are updated due to factors such as technological iteration. Therefore, it is necessary to upgrade the algorithm on devices that have been deployed and used to improve the reliability of the algorithm. Usually, when performing algorithm upgrade, face images can be directly stored for feature upgrade of the images. However, this algorithm upgrade method is applicable to devices with sufficient storage resources. For end-side devices with limited storage resources (such as door locks, door viewers, etc.), storing additional face images means increasing storage resources, which will increase the cost of the device and is not conducive to the popularization of the device. In addition, when the algorithm is upgraded, users are also required to cooperate to re-register information, making the entire upgrade process complicated and affecting the user experience. Summary of the Invention

[0003] To at least solve one or more of the above-mentioned technical problems, this disclosure proposes a highly practical and support-for-seamless-upgrade solution for updating a biometric recognition model in multiple aspects.

[0004] In a first aspect of the embodiments of this disclosure, a method for updating a biometric recognition model is provided, including: in response to obtaining a requirement to update an old biometric recognition model to a new biometric recognition model, performing style transfer processing on the registered biometrics corresponding to the old biometric recognition model to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model; in response to obtaining a biometric recognition requirement for an image to be recognized, using the new biometric recognition model to process the image to be recognized to obtain new model features; and selectively replacing the new learning features in the new learning feature set with the new model features to implement the update of the registered features of the new biometric recognition model.

[0005] In some embodiments, selectively replacing the new learning features in the new learning feature set with the new model features includes: in response to obtaining a new learning feature that matches the new model feature from the new learning feature set, replacing the obtained new learning feature with the new model feature.

[0006] In some embodiments, where the newly learned features in the newly learned feature set are configured with a first identifier, for the newly learned features configured with the first identifier, the following steps are performed: obtaining new model features corresponding to at least one frame of the image to be recognized; detecting whether there are newly learned features configured with the first identifier in the newly learned feature set that match the new model features corresponding to the at least one frame of the image; and in response to there being newly learned features configured with the first identifier that match the new model features corresponding to the at least one frame of the image, determining that the newly learned features matching the new model features are obtained from the newly learned feature set.

[0007] In some embodiments, the method further includes: performing a marking process on the new model features replacing the newly learned features by using a second identifier, where the second identifier is different from the first identifier.

[0008] In some embodiments, performing a style transfer process on the registered biometric features corresponding to the old biometric recognition model includes: obtaining a pre-trained feature style transfer model; and using the feature style transfer model to perform a style transfer process on the registered biometric features to obtain the newly learned feature set.

[0009] In some embodiments, where the feature style transfer model is trained through the following steps: obtaining first training data obtained by performing biometric feature extraction based on the old biometric recognition model; obtaining second training data obtained by performing biometric feature extraction based on the new biometric recognition model; and based on the first training data and the second training data, performing style transfer training on the base model of the feature style transfer model to obtain the feature style transfer model, where the trained feature style transfer model supports style transfer between the first training data and the second training data.

[0010] In some embodiments, based on the first training data and the second training data, performing style transfer training on the base model of the feature style transfer model includes: inputting the first training data into the base model and obtaining intermediate training data output by the last layer network of the base model; calculating a loss function between the intermediate training data and the second training data; and performing optimization training on the base model based on the loss function.

[0011] In some embodiments, where the biometric recognition requirement includes a face recognition requirement, calculating the loss function between the intermediate training data and the second training data includes: calculating the face recognition loss function ArcFace loss between the intermediate training data and the second training data.

[0012] In a second aspect of the embodiments of the present disclosure, there is provided an electronic device, further comprising: a processor; and a memory storing computer instructions for updating a biometric recognition model, which, when run by the processor, cause the electronic device to execute the methods described in multiple embodiments above and below.

[0013] In a third aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium including program instructions for updating a biometric recognition model, which, when executed by a processor, cause the implementation of the methods described in multiple embodiments above and below.

[0014] Through the solution for updating the biometric recognition model provided as above, the embodiments of the present disclosure can utilize the style transfer processing of the registered biometrics corresponding to the old biometric recognition model to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model, and in the process of recognizing the image to be recognized, selectively replace the new learning features in the new learning feature set with the new model features to achieve the update of the registered features of the new biometric recognition model. It can be seen that in the process of model update, the solution of the present disclosure can utilize the learning of the style of the old or registered biometrics to obtain a new learning feature set, and use the new learning feature set to achieve the initial update of the registered features of the new biometric recognition model. Thus, using the new learning feature set with lower storage resource requirements to achieve the initial update of the model, on the one hand, compared with the traditional technology of updating the algorithm using pictures, it can effectively save storage resources and reduce the requirements for the storage resources of the device, thereby effectively expanding the application scenarios and having stronger practicability. On the other hand, the entire initial upgrade process does not require excessive user participation (for example, there is no need for the user to re-register), realizing a seamless upgrade and improving the user experience.

[0015] In addition, in the subsequent image recognition process, selectively replace the new learning features in the new learning feature set with the new model features, so as to achieve the final update of the registered features of the new biometric recognition model. Thus, the possible accuracy loss between the new learning feature set and the new model features can be reduced, effectively reducing the accuracy loss in the model update process.

[0016] Further, in some embodiments, a pre-trained feature style transfer model can be used to implement the style transfer processing between features, and it only needs to run the feature style transfer model once when the new biometric recognition model is deployed, and there is no need to run it subsequently, so that the resource occupancy is small or does not occupy the device resources during the subsequent user application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present disclosure will become readily understandable. In the drawings, several embodiments of the present disclosure are shown in an exemplary but non-limiting manner, and the same or corresponding reference numerals denote the same or corresponding parts, wherein:

[0018] Figure 1 A schematic flow chart of a method for updating a biometric recognition model according to an embodiment of the present disclosure is shown;

[0019] Figure 2 A schematic flow chart of a method for updating a biometric recognition model according to another embodiment of the present disclosure is shown;

[0020] Figure 3 A schematic flow chart of a method for updating a biometric recognition model according to still another embodiment of the present disclosure is shown;

[0021] Figure 4 A schematic flow chart of a method for training a feature style transfer model according to an embodiment of the present disclosure is shown;

[0022] Figure 5 A schematic flow chart of a method for training a feature style transfer model according to another embodiment of the present disclosure is shown; and

[0023] Figure 6 A schematic structural diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Embodiments

[0024] Next, the technical solutions in the embodiments of the present disclosure will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative efforts fall within the scope of protection of the present disclosure.

[0025] It should be understood that the terms "including" and "comprising" used in the specification and claims of the present disclosure indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0026] It should also be understood that the terms used in this disclosure specification are only for the purpose of describing specific embodiments and are not intended to limit the disclosure. As used in this disclosure specification and the claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms. It should also be further understood that the term "and / or" used in this disclosure specification and the claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0027] As used in this specification and the claims, the term "if" can be interpreted as "when", "once", "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if [the described condition or event] is detected" can be interpreted as meaning "once determined", "in response to determining", "once [the described condition or event] is detected" or "in response to detecting [the described condition or event]" depending on the context.

[0028] Devices that generally support biometric recognition such as face recognition need to upgrade the recognition algorithm models thereon to improve the reliability of the algorithms. For some devices with sufficient storage resources, pictures can be directly stored for algorithm upgrade. For some end - side devices with limited storage resources (such as door locks, door viewers for identification, etc.), storing additional pictures means increasing storage resources, which will increase the cost of the device and is not conducive to the popularization of the device. Therefore, in related technologies, algorithm upgrade can be achieved by registering and saving biometric features.

[0029] The inventors found that for different algorithm models, features cannot be directly reused. In the traditional algorithm upgrade process, it is necessary to re - register the face database, which overly relies on the cooperation of users, resulting in a poor user experience. Especially in some application scenarios of algorithm upgrade (such as face recognition used at home), if the user does not re - register in time, it usually leads to the failure of face recognition, making the user experience poor and inconvenient to use.

[0030] Based on this, the inventors have found through research that due to factors such as different styles between features, different models cannot reuse features. In this regard, the existing or registered features and the feature - style transfer technology can be used to obtain features related to the biometric feature style corresponding to the new model. Specifically, when upgrading the biometric recognition model, the registered feature style corresponding to the old biometric recognition model can be transferred to the feature space of the new biometric recognition model to achieve the style transfer between features, and in the subsequent biometric recognition process, online update can be performed on the features after the style transfer. Thus, the seamless upgrade of the biometric recognition model is realized.

[0031] The specific embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0032] Figure 1 FIG. 5 shows a flowchart of a method 100 for updating a biometric recognition model according to an embodiment of the present disclosure. It should be noted that the specific type of the biometric recognition model in the present disclosure is not limited. For example, the biometric recognition model in the present disclosure may include a face feature recognition model, a palmprint feature recognition model, a palm vein feature recognition model, or other models that need to rely on the update of biometric data to achieve model upgrade or update.

[0033] Specifically, it can be implemented through Figure 1 the method 100 in FIG. 5 to perform the update process for the above-mentioned biometric recognition model.

[0034] As Figure 1 shown, at step S101, in response to obtaining the requirement to update the old biometric recognition model to a new biometric recognition model, the registered biometrics corresponding to the aforementioned old biometric recognition model can be subjected to style transfer processing to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model.

[0035] In this embodiment, the old and new biometric recognition models can be understood as different updated versions of the biometric model, or the old biometric recognition model can also be understood as the biometric recognition model already installed in the device, and the new biometric recognition model is understood as the new model used to replace the installed biometric recognition model.

[0036] In practical applications, there are various ways to determine whether there is a need to update the aforementioned old and new biometric recognition models. For example, in some embodiments, it can be detected whether a new biometric recognition model used to replace the old biometric recognition model is installed in the device. If the new biometric recognition model is detected to be installed, it is determined that there is a need to update the old biometric recognition model to the new biometric recognition model. For another example, it can be detected whether there is a download requirement for the new biometric recognition model. If there is such a download requirement, the new biometric recognition model is downloaded and installed, and in response to the completion of the installation of the new biometric recognition model, it is determined that there is a need to update the old biometric recognition model to the new biometric recognition model. For still another example, in response to determining that a new biometric recognition model is installed in the device, it is detected whether a specified operation of the user for the new biometric recognition model is obtained (such as a click, a press, or other forms of human-computer interaction operations, and the specified operation can be set and adjusted in combination with specific interaction requirements). If the specified operation is obtained, it is determined that there is a need to update the old biometric recognition model to the new biometric recognition model. It should be noted that the description here of whether there is a need to update the model between the old and new biometric recognition models is only an exemplary illustration, and the solution of the present disclosure is not limited thereto.

[0037] When determining the need to update the old biometric recognition model to the new biometric recognition model, style transfer processing can be performed on the registered biometrics corresponding to the aforementioned old biometric recognition model to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model. As mentioned above, the styles of different features are different, and the existing or registered features and the style transfer technology between features can be used to obtain features related to the biometric style corresponding to the new model. Specifically, in this embodiment, the style transfer technology between features can be used to perform style transfer processing on the registered biometrics corresponding to the old biometric recognition model to obtain a new learning feature set, and the new learning features in the new learning feature set are related to the biometric style corresponding to the new biometric recognition model.

[0038] It should be understood that pictures have much higher requirements for storage resources than feature sets. In this embodiment, the new learning feature set with lower requirements for storage resources is used to realize the preliminary update of the registered features of the new biometric recognition model. On the one hand, compared with the traditional technology of using pictures for algorithm update, it can effectively save storage resources and reduce the requirements for the storage resources of the device, thereby effectively expanding the application scenarios and having stronger practicability. On the other hand, the entire preliminary upgrade process does not require the user to re-register the features, realizing a seamless upgrade and improving the user experience.

[0039] Further, at step S102, in response to obtaining a biometric recognition requirement for the image to be recognized, the new biometric recognition model can be used to process the image to be recognized to obtain new model features.

[0040] In this embodiment, the foregoing biometric recognition requirements may include face recognition requirements, palmprint recognition requirements, or palm vein recognition requirements, etc. In practical applications, there are various ways to determine whether the biometric recognition requirement is obtained. For example, an image acquisition device (such as a camera) is provided on the device. When the image to be recognized is obtained based on the camera, it can be determined that the biometric recognition requirement for the image to be recognized is obtained. It should be noted that the specific acquisition process of the biometric recognition requirement in the present disclosure is not limited, and the triggering mechanism of the biometric recognition requirement can be set according to application requirements.

[0041] At step S103, the new learned features in the new learned feature set can be selectively replaced with the new model features to update the registered features of the new biometric recognition model. After obtaining the new model features, the new model features can be compared with the new learned features in the new learned feature set, and then the new learned features in the new learned feature set can be selectively replaced with the new model features according to the comparison result to update the registered features of the new biometric recognition model.

[0042] Considering that there may be a precision loss between the new learned feature set and the new model features, the solution of the present disclosure selectively replaces the new learned features in the new learned feature set in combination with the new model features during the subsequent image recognition process, thereby realizing the final update of the registered features of the new biometric recognition model. Thus, the possible precision loss between the new learned feature set and the new model features can be reduced, and the precision loss during the algorithm update process can be effectively reduced.

[0043] Based on this, in the process of updating the biometric recognition model, the solution of the present disclosure can not only realize the seamless upgrade of the model, but also effectively reduce the precision loss during the model update process, so as to improve the recognition accuracy of the model, and has stronger practicability and better market prospects.

[0044] Figure 2 FIG. shows a schematic flowchart of a method 200 for updating a biometric recognition model according to another embodiment of the present disclosure. It should be noted that Figure 2 The method 200 can be understood as a further limitation or extension of Figure 1 The method 100 in Figure 1 Therefore, the relevant descriptions in the foregoing in combination with

[0046] Such as Figure 2As shown, at step S201, in response to obtaining the requirement to update the old biometric recognition model to a new biometric recognition model, a pre-trained feature style transfer model is obtained. In this embodiment, for the specific details of the old and new biometric recognition models and the update requirements between the models, reference may be made to the relevant descriptions in the foregoing text. Figure 1 They will not be elaborated here.

[0047] The foregoing feature style transfer model can be pre-trained and support style transfer between features. In practical applications, the training of the feature style transfer model can be achieved through supervised training and other methods. The trained feature style transfer model can be pre-stored locally or on the server side. When the requirement to update the old biometric recognition model to a new biometric recognition model is obtained, the feature style transfer model can be retrieved from the local side or downloaded in real time from the server side. It should be noted that the description of the process of obtaining the feature style transfer model here is only an exemplary illustration, and the solution disclosed in this disclosure is not limited thereto.

[0048] At step S202, the foregoing feature style transfer model can be used to perform style transfer processing on the registered biometrics to obtain a new learning feature set. Specifically, after obtaining the feature style transfer model, the feature style transfer model can be used to perform style transfer processing on the registered biometrics corresponding to the old biometric recognition model to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model. For example, the registered biometrics corresponding to the old biometric recognition model are used as the input of the feature style transfer model, and the feature style transfer model can output new learning features related to the biometric style corresponding to the new biometric recognition model, and a new learning feature set is formed based on the new learning features output by the feature style transfer model.

[0049] In practical applications, when the feature style transfer model is deployed to the device, or when both the new biometric recognition model and the feature style transfer model are deployed to the device, the feature style transfer model can be run once to obtain the new learning feature set, and there is no need to run the feature style transfer model again later, so that the resource occupancy is small or the device resources during subsequent user applications are not occupied.

[0050] Furthermore, in response to obtaining the biometric recognition requirement for the image to be recognized, the new biometric recognition model can be used to process the image to be recognized to obtain new model features. Then, in response to obtaining a new learning feature that matches the new model feature from the new learning feature set, the obtained new learning feature can be replaced with the new model feature.

[0051] Specifically, in some embodiments, the new learning features in the aforementioned new learning feature set are configured with a first identifier, and a replacement operation between features can be performed on the new learning features configured with the first identifier in the new learning feature set. For specific reference, please refer to Figure 2 Steps S203 to S205 in

[0052] Among them, at step S203, in response to obtaining a biometric recognition requirement for the image to be recognized, at least one frame of image corresponding new model features of the image to be recognized can be obtained. For example, when the image to be recognized is collected by an image acquisition device on the device, it can be determined that a biometric recognition requirement for the image to be recognized is obtained. At this time, the new biometric recognition model can be used to process at least one frame of image of the image to be recognized to obtain at least one frame of image corresponding new model features of the image to be recognized. It should be noted that the number of images participating in feature extraction is not limited here. For example, only one frame of image of the image to be recognized can be processed, or multiple frames of images (including consecutive multiple frames of images or non-consecutive multiple frames of images) of the image to be recognized can be processed.

[0053] At step S204, it is detected whether there is a new learning feature configured with the first identifier in the new learning feature set that matches the new model features corresponding to at least one frame of image. In some embodiments, when only one frame of image is processed, a new learning feature configured with the first identifier that matches the new model features corresponding to the single frame of image can be searched for in the new learning feature set. In other embodiments, when multiple frames of images are processed, a new learning feature configured with the first identifier that matches the new model features corresponding to any one of the frames of images can be searched for in the new learning feature set, and then the new model features corresponding to the remaining other frames of images are used to perform a matching confirmation with this new learning feature again.

[0054] At step S205, in response to the existence of a new learning feature configured with the first identifier that matches the new model features corresponding to at least one frame of image, it is determined that the new learning feature that matches the new model features is obtained from the new learning feature set. For example, when it is necessary to use the new model features corresponding to a single frame of image for feature matching confirmation, the new learning feature configured with the first identifier that matches the new model features corresponding to the single frame of image can be determined as the new learning feature to be updated obtained. Another example is that when it is necessary to use the new model features corresponding to multiple frames of images for feature matching confirmation, the new learning feature configured with the first identifier that matches all the new model features corresponding to the multiple frames of images can be determined as the new learning feature to be updated obtained. By performing multiple feature matching confirmations through the new model features corresponding to multiple frames of images, the accuracy of feature matching can be ensured.

[0055] Then, at step S206, the newly obtained learning feature can be replaced with a new model feature. After obtaining the new learning feature to be updated, the new model feature can be used to replace the new learning feature. For example, the new learning feature can be deleted from the new learning feature set, and the new model feature replacing the new learning feature can be saved.

[0056] Thus, the solution of this embodiment can use the pre-trained feature style transfer model to achieve style transfer between old and new features, and support the initial update of the new biometric recognition model based on the new learning feature set. In addition, during the process of feature recognition using the new biometric recognition model, the obtained new model feature can be used to update the new learning feature set online to complete the final update of the new biometric recognition model. It can be seen that the update process of the entire biometric recognition model does not require the user to register repeatedly, reducing the user's participation to achieve a seamless upgrade of the model.

[0057] It can be understood that during the process of feature recognition using the new biometric recognition model, on the one hand, the new model feature can be used to match with the new learning feature. If there is a new learning feature that matches the new model feature, it is determined that the biometric recognition of the image to be recognized passes. At the same time, when it is determined that there is a new learning feature that matches the new model feature, the new learning feature can be replaced with the new model feature. Thus, it can not only ensure the normal use of the new model, but also reduce the accuracy loss between the new learning feature and the new model feature, improving the recognition accuracy of the new model.

[0058] In addition, in actual application scenarios (such as office area access control, etc.), since the users change relatively frequently, some new learning features in the new learning feature set may not be used. This embodiment uses the obtained new model feature to update the new learning feature set online, which can specifically update the features that need to be used, avoiding meaningless update operations on some unused features, and being more in line with actual needs. Of course, in actual applications, if all the new learning features in the new learning feature set will be used, then the solution based on this embodiment can ultimately also achieve the update of all the new learning features in the new learning feature set.

[0059] Furthermore, in some embodiments, the new model feature replacing the new learning feature can also be marked using a second identifier. The second identifier is different from the first identifier. Thus, when the image to be recognized of the same user is recognized again, the new model feature with the second identifier can be used to perform normal feature recognition matching with the biometric feature extracted from the image to be recognized, without the need for feature update.

[0060] Further, in some embodiments, there may be some new users. At this time, the new learning features in the new learning feature set may not be able to match the new model features corresponding to the images of the new users. At this time, a prompt can be output to prompt the new users to register information.

[0061] Figure 3 FIG. 4 shows a schematic flowchart of a method 300 for updating a biometric recognition model according to another embodiment of the present disclosure. It should be noted that Figure 3 the method 300 can be understood as a Figure 1 specific technical implementation of the method 100 or Figure 2 the method 200. Therefore, the relevant descriptions in the foregoing in combination with Figure 1 and Figure 2 also apply to the following.

[0062] Specifically, as Figure 3 shown, at step S301, a feature style transfer model can be pre-trained and the feature style transfer model can be deployed. When updating the biometric recognition model, the feature style transfer model is started, and all the registered old features are learned as new learning features. In some embodiments, the feature style transfer model can be started only once when updating the new biometric recognition model to complete the style transfer processing between features, and there is no need to run the feature style transfer model again later, so that the device resources are less occupied or the device resources during subsequent user applications are not occupied.

[0063] In practical applications, the feature style transfer model can be pre-trained by means of supervised training and the like. Specifically, the following in combination with Figure 4 and Figure 5 illustrates the specific training process of the foregoing feature style transfer model.

[0064] Figure 4 FIG. 5 shows a schematic flowchart of a method 400 for training a feature style transfer model according to an embodiment of the present disclosure. In this embodiment, the training process of the feature style transfer model can include a training data acquisition stage and a model training stage. Among them, in the training data acquisition stage, the first training data obtained by using the old biometric recognition model and the second training data obtained by using the new biometric recognition model can be used to form a training data pair. In the model training stage, the foregoing training data pair can be input into the feature style transfer model to train the model, so that the style of the first training data can be transferred to the style of the second training data. The new learning feature style output by the trained feature style transfer model is similar or analogous to the feature style of the new biometric recognition model.

[0065] Specifically, as Figure 4As shown, at step S401, first training data obtained by performing biometric feature extraction based on an old biometric recognition model can be acquired. In this embodiment, some common data collection techniques can be used to collect some publicly available biometric data, and the old biometric recognition model can be used to perform feature extraction on this biometric data (such as face images, etc.) to obtain first biometric data, and then first training data can be obtained based on the first biometric data.

[0066] At step S402, second training data obtained by performing biometric feature extraction based on a new biometric recognition model can be acquired. For example, the new biometric recognition model can be used to perform feature extraction on the aforementioned collected biometric data such as face images to obtain second biometric data. Then, second training data can be obtained based on the second biometric data. It should be noted that the specific execution order of steps S401 and S402 in the present disclosure is not limited. For example, S401 and S402 can be executed sequentially or synchronously.

[0067] At step S403, based on the first training data and the second training data, style transfer training can be performed on the base model of the feature style transfer model to obtain the feature style transfer model. Among them, the trained feature style transfer model supports style transfer between the aforementioned first training data and the second training data. Specifically, a training data pair can be constructed using the first training data and the second training data, and the constructed training data pair can be input into the feature style transfer model to perform style transfer training on the base model of the feature style transfer model.

[0068] In some embodiments, the first training data in the training data pair can be used as the model input, and the second training data in the training data pair can be used as the label to implement supervised training of the feature style transfer model based on the training data pair.

[0069] Figure 5 FIG. shows a schematic flowchart of a method 500 for training a feature style transfer model according to another embodiment of the present disclosure. It should be noted that method 500 can be understood as a specific technical implementation of method 400. Therefore, the relevant details described above in conjunction with Figure 4 are equally applicable to the following.

[0070] In this embodiment, the first training data can be input into the base model of the feature style transfer model, and the intermediate training data output by the last layer network of the base model can be acquired. Then, the loss function between the intermediate training data and the second training data is calculated. Then, the base model is optimized and trained based on the aforementioned loss function.

[0071] Specifically, as Figure 5As shown, the base network of the feature style transfer model in this embodiment may include a combination of multiple sets of fully connected layers (Fully Connected Layer, abbreviated as FC) + activation layers (such as Parametric Rectified Linear Unit, abbreviated as PRelu). At step S501, the first training data can be used as the input of the base model. After being processed by multiple sets of fully connected layers + activation layers, the feature vector output by the last layer PRelu of the base network (i.e., the aforementioned intermediate training data) is obtained. At step S502, the second training data serving as the label is acquired. Then, at step S503, the loss function between the intermediate training data output by the last layer PRelu of the base network and the second training data serving as the label is calculated, and forward propagation calculation is performed based on this loss function to finally achieve the optimization training of the base model. It should be noted that the description of the architecture of the base network here is only an exemplary illustration, and the solution of this disclosure does not limit this, and it can be specifically set and adjusted according to requirements.

[0072] In some embodiments, if the feature style transfer model is for the style transfer of face features, the face recognition loss function ArcFace loss between the intermediate training data and the second training data can be calculated. Of course, according to different biometric recognition requirements (such as face recognition, palmprint recognition, palm vein recognition, etc.), other types of loss functions between the intermediate training data and the second training data can be used to achieve the optimization training of the base model.

[0073] After training the feature style transfer model using the Figure 4 and Figure 5 shown method, return Figure 3 In step S301, the feature style transfer model and the new biometric recognition model can be deployed to the device. In a specific application, in response to the need to update the old biometric recognition model to a new biometric recognition model, the feature style transfer model is started, and all registered old biometric features (such as the registered biometric features corresponding to the old biometric recognition model) are input into the feature style transfer model for style transfer, and then new learning features related to the biometric style corresponding to the new biometric recognition model are output based on the feature style transfer model.

[0074] In step S302, a new learning feature set can be constructed using the new learning features output by the feature style transfer model.

[0075] In step S303, during the biometric recognition process, a to-be-recognized image can be acquired. For example, during face recognition, a face is captured to obtain the to-be-recognized face image.

[0076] At step S304, the new biometric recognition model can be used to process the image to be recognized. After all the registered old features are updated to new learned features, during the recognition process, the new biometric recognition model deployed in the device can be used to extract features from the image to be recognized.

[0077] At step S305, the new model features corresponding to the image to be recognized can be obtained.

[0078] At step S306, the new learned features in the new learned feature set are compared with the new model features.

[0079] In some embodiments, during the model update and conversion process, the new learned features (such as face features, etc.) in the new learned feature set are marked. For example, the new learned features can be marked as 0 (i.e., the aforementioned first identifier). When performing biometric recognition and comparison, if it is found that the current new model features meet the comparison conditions of the new learned features (such as the similarity between the new model features and the new learned features is greater than a predetermined value, etc.), a secondary confirmation will be performed. If the new model features corresponding to multiple frames (such as 2 consecutive frames) of images all meet the comparison conditions of the new learned features, then the new model features are used to replace the new learned features in the new learned feature set, and the marking information is updated to 1 (i.e., the second identifier). Subsequently, for the new model features marked as 1, normal biometric recognition can be performed using the new model features marked as 1, and biometric updates are no longer performed. It should be noted that the comparison conditions between the new model features and the new learned features in the solution of the present disclosure are not limited, and feature comparison can be achieved through various feature analysis techniques. In addition, the first identifier and the second identifier can be in digital form, or in the form of text or other identification information.

[0080] At step S307, feature recognition is completed, and the corresponding new learned features are replaced with new model features. Specifically, if there are new learned features in the new learned feature set that match the new model features, it means that the features corresponding to the image to be recognized already exist in the new learned feature library, and the matching new learned features are replaced with new model features to reduce the accuracy loss between the new model features and the new learned features.

[0081] Thus, the solution of this embodiment can reuse the saved biometric features, and the update of the biometric recognition model can be achieved without the user having to register repeatedly. In addition, the pre-trained feature style transfer model can perform style transfer on the features extracted by two different biometric recognition algorithm models, that is, the features of a biometric recognition model can be mapped to the feature space of another biometric recognition model after being processed by the feature style transfer model, so as to achieve style conversion between features. In addition, by deploying the feature style transfer model on the device side, when the device recognition algorithm or model is upgraded, a biometric style transfer is performed once, and the saved registered biometric style can be transferred to the biometric space of the new model. During subsequent recognition, online update and coverage are performed on the new learned feature set after style transfer. Thus, not only can

[0082] the recognition accuracy be improved, but also the feature style transfer model is only started once, which occupies little device resources or does not occupy the device resources during subsequent user applications.

[0083] After introducing the method of the exemplary embodiment of this disclosure, next, reference is made to Figure 6 describe the electronic device of the exemplary embodiment of this disclosure.

[0084] Figure 6 A schematic block diagram of an electronic device 600 according to an embodiment of the present disclosure is schematically shown.

[0085] As Figure 6 shown, the electronic device 600 may include a processor 601 and a memory 602. The memory 602 stores computer instructions for updating the biometric recognition model. When the computer instructions are run by the processor 601, the electronic device 600 is caused to execute in accordance with the foregoing in combination with Figures 1 to 5The described method. For example, in some embodiments, during the model update process, the electronic device 600 can utilize the learning of the old or registered biometric styles to obtain a new learned feature set, and use the new learned feature set to implement the preliminary update of the registered features of the new biometric recognition model. Thus, using the new learned feature set with lower storage resource requirements to implement the preliminary update of the model, on the one hand, compared with the traditional technology of updating the algorithm using pictures, it can effectively save storage resources and reduce the requirements for the device's storage resources, thereby effectively expanding the application scenarios and having stronger practicality. On the other hand, the entire preliminary upgrade process does not require excessive user participation (for example, there is no need for the user to re-register), realizing a seamless upgrade and enhancing the user's experience. Additionally, during the subsequent image recognition process, selectively replace the new learned features in the new learned feature set in combination with the new model features, thereby realizing the final update of the registered features of the new biometric recognition model. Thus, it is possible to reduce the possible accuracy loss between the new learned feature set and the new model features, thereby effectively reducing the accuracy loss during the algorithm update process.

[0086] In addition, the present disclosure also provides a computer-readable storage medium storing program instructions configured to execute when running Figures 1 to 5 the method for updating a biometric recognition model shown in any one of the figures in

[0087] Specifically, in this embodiment, the above storage medium may include, but is not limited to: USB flash drives, read-only memories (ROM), random access memories (RAM), external hard drives, magnetic disks, or optical discs, etc., various media that can store computer programs.

[0088] It should be noted that although several devices or sub-devices of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of the two or more devices described above can be embodied in one device. Conversely, the features and functions of one device described above can be further divided and embodied by multiple devices.

[0089] Although the spirit and principles of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of each aspect does not mean that the features in these aspects cannot be combined for benefits. This division is only for the convenience of expression. The present disclosure aims to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. The scope of the appended claims is interpreted in the broadest sense to encompass all such modifications and equivalent structures and functions.

Claims

1. A method for updating a biometric recognition model, characterized in that, Including: In response to obtaining a requirement to update an old biometric recognition model to a new biometric recognition model, performing a style transfer process on the registered biometrics corresponding to the old biometric recognition model to obtain a new learning feature set related to the biometric style corresponding to the new biometric recognition model; In response to obtaining a biometric recognition requirement for an image to be recognized, using the new biometric recognition model to process the image to be recognized to obtain new model features; And Selectively replacing the new learning features in the new learning feature set with the new model features to update the registered features of the new biometric recognition model.

2. The method according to claim 1, characterized in that, Selectively replacing the new learning features in the new learning feature set with the new model features includes: In response to obtaining a new learning feature that matches the new model feature from the new learning feature set, replacing the obtained new learning feature with the new model feature.

3. The method according to claim 2, characterized in that, Wherein the new learning features in the new learning feature set are configured with a first identifier. For the new learning features configured with the first identifier, the following steps are performed: Obtaining new model features corresponding to at least one frame of the image to be recognized; Detecting whether there is a new learning feature configured with the first identifier in the new learning feature set that matches the new model features corresponding to the at least one frame of the image; And In response to there being a new learning feature configured with the first identifier that matches the new model features corresponding to the at least one frame of the image, determining that a new learning feature that matches the new model feature is obtained from the new learning feature set.

4. The method according to claim 3, characterized in that, The method further includes: Marking the new model features that replace the new learning features with a second identifier, where the second identifier is different from the first identifier.

5. The method according to any one of claims 1 to 4, characterized in that, Performing a style transfer process on the registered biometrics corresponding to the old biometric recognition model includes: Obtaining a pre-trained feature style transfer model; and Using the feature style transfer model to perform a style transfer process on the registered biometrics to obtain the new learning feature set.

6. The method according to claim 5, characterized in that, Wherein the feature style transfer model is trained through the following steps: Obtaining first training data obtained by performing biometric extraction based on the old biometric recognition model; Obtaining second training data obtained by performing biometric extraction based on the new biometric recognition model; And Based on the first training data and the second training data, performing style transfer training on the base model of the feature style transfer model to obtain the feature style transfer model, where the trained feature style transfer model supports style transfer between the first training data and the second training data.

7. The method according to claim 6, characterized in that, Performing style transfer training on the base model of the feature style transfer model based on the first training data and the second training data includes: Inputting the first training data into the base model and obtaining intermediate training data output by the last layer network of the base model; Calculating a loss function between the intermediate training data and the second training data; and Based on the loss function, performing optimization training on the base model.

8. The method according to claim 7, characterized in that, Wherein the biometric recognition requirement includes a face recognition requirement, and calculating a loss function between the intermediate training data and the second training data includes: Calculating the face recognition loss function ArcFace loss between the intermediate training data and the second training data.

9. An electronic device, characterized in that, It further includes: A processor; And A memory storing computer instructions for updating a biometric recognition model, and when the computer instructions are run by the processor, the electronic device is caused to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that,Including program instructions for updating a biometric recognition model, and when the program instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images