JP2022116231A - Training method of organism detection model, device, electronic apparatus and storage medium - Google Patents

Abstract

To provide a training method of an organism detection mold, a device, an electronic apparatus and a storage medium.SOLUTION: A method includes the steps of: constructing a training set for training an organism detection mold, and a test set; training a preset organism detection model on the basis of the training set, and acquiring a first organism detection model; training the first organism detection mold on the basis of the test set, and creating a test result; analyzing the test set on the basis of the test result, and acquiring first sample data; and updating the training set and the test set on the basis of the first sample data, and further training the organism detection model.EFFECT: Model training efficiency can be improved by introducing a concept of sample mining from a data layer, and effectively reducing a sample having a large amount of redundancy and having no instruction significance.SELECTED DRAWING: Figure 1

Description

The present disclosure relates to the field of artificial intelligence technology, specifically to the field of computer vision and deep learning technology, and can be applied to scenes such as face recognition.

Face biometrics detection, that is, distinguishing whether an image is a photograph of a real person or not, is a basic module of a face recognition system and guarantees the safety of the face recognition system. Face biometric detection methods using deep learning technology are currently the mainstream in this field, and their accuracy is greatly improved compared to conventional methods. However, in practical applications, due to the diversification of face attack samples, various attack schemes are constantly emerging, and the optimized model has very limited ability to resist new attacks that have not yet been seen. High cost and low efficiency.

The present disclosure provides methods, apparatus, electronics and storage media for training liveness detection models.

According to one aspect of the present disclosure, constructing a training set and a test set for training a liveness detection model; training a preset liveness detection model based on the training set; training a first liveness detection model based on the test set to generate test results; and analyzing the test set based on the test results to obtain first sample data. and further training the liveness detection model by updating a training set and a test set based on the first sample data.

According to another aspect of the present disclosure, a building module for building a training set and a test set for training a liveness detection model; training a preconfigured liveness detection model based on the training set; a training module for obtaining a first liveness detection model, training the first liveness detection model based on the test set, and generating test results; analyzing the test set based on the test results; Training a liveness detection model, including a sample acquisition module for acquiring sample data and an update module for further training the liveness detection model by updating a training set and a test set based on first sample data. Provide equipment.

According to another aspect of the present disclosure, an electronic device is provided, said electronic device including at least one processor and memory communicatively coupled to said at least one processor, said memory comprising: and executable by the at least one processor, said instructions being executed by said at least one processor to cause said at least one processor to perform the method described above.

According to another aspect of the disclosure, there is provided a non-transitory computer-readable storage medium having computer instructions stored thereon for causing a computer to perform the method described above.

According to another aspect of the disclosure, there is provided a computer program product comprising a computer program that, when executed by a processor, implements the method described above.

The present disclosure introduces the idea of sample mining from the data layer, which can effectively reduce a large amount of redundant, non-instructive samples and improve model training efficiency.

It should be understood that the content described in this part is not intended to show the gist or important features of the embodiments of the present disclosure, and is not intended to limit the protection scope of the present disclosure. is. Other features of the present disclosure will become easier to understand with the following specification.

The drawings are for a better understanding of the invention and do not constitute limitations on the disclosure.
FIG. 2 illustrates a method for training a liveness detection model according to a first embodiment of the present invention; FIG. 5 illustrates a method for training a liveness detection model according to a second embodiment of the present invention; FIG. 10 is a diagram showing the configuration of a liveness detection model training device according to a third embodiment of the present disclosure; 1 is a block diagram of an electronic device for implementing a liveness detection model training method according to an embodiment of the present disclosure; FIG.

Illustrative embodiments of the disclosure will now be described in conjunction with the drawings, and various details in the embodiments of the disclosure contained therein are for the sake of understanding and should be considered as illustrative only. be. Accordingly, it should be understood by those skilled in the art that various changes and modifications can be made to the examples described herein without departing from the scope and spirit of this disclosure. Similarly, for the sake of clarity and brevity, the following description omits descriptions of well-known functions and constructions.

In an embodiment of the present disclosure, new data refers to newly appearing new face attack sample data after the liveness detection model is optimized. Raw data means the data from which the liveness detection model was originally trained.

The present disclosure provides a method for training a liveness detection model based on hard sample mining, retraining the model using a mixture of new and original data to solve the accuracy loss due to the model's catastrophic forgetting problem, At the same time, based on the idea of hard sample mining, by improving the ratio of hard samples in training samples, we can reduce redundant samples, improve model training efficiency, and make the model pay more attention to hard samples, so that the model can be inspected. improve performance.

FIG. 1 is a diagram illustrating a liveness detection model training method according to a first embodiment of the present disclosure. As shown in FIG. 1, the liveness detection model training method according to the first embodiment of the present disclosure includes the following steps.

Step 101 builds a training set and a test set for training a liveness detection model.

Constructing a training set and a test set for training the liveness detection model includes constructing a training set and a test set based on a mixed data set, the mixed data set including new data and original data. .

By constructing a mixed dataset with new and original data, and building training and test sets based on the mixed dataset, the accuracy loss due to the catastrophic forgetting problem of the model can be resolved.

Constructing a training set and a test set based on the mixed data set includes: performing random sampling on the mixed data set; configuring the sampled data as a training set; as a test set.

By constructing the initial training set and initial test set by performing random sampling on the mixed dataset, the training effect on the model can be better realized.

Performing random sampling on the mixed data set includes performing random sampling on the mixed data set based on an initial sampling rate of preset hyperparameters.

By setting the initial sampling rate of the random sampling hyperparameters, the random sampling index can be specifically limited.

Here, in this embodiment, the initial sampling rate of the hyperparameter takes a value greater than 0% and less than 50%. Of course, in practical applications, other settings may be made to the initial sampling rate of the hyperparameters as needed, and the present disclosure is not limited thereto.

By selecting the value of the initial sampling rate of the hyperparameters, it provides guidelines for practical applications, provides specific implementation methods for embodiments of the present disclosure, and facilitates selection in practical applications.

In step 102, a preset liveness detection model is trained based on the training set to obtain a first liveness detection model.

At step 103, a first liveness detection model is trained based on the test set to generate test results.

In step 104, the test set is analyzed based on the test results to obtain first sample data.

In an embodiment of the present disclosure, analyzing the test set based on the test results to obtain first sample data comprises scoring data in the test set based on the test results to obtain a prediction score; Sorting the data in the test set based on the prediction score, and determining the data whose prediction score meets a predetermined threshold as first sample data.

Therefore, the embodiment of the present invention implements a specific implementation method for obtaining the first sample data, implements mining of the first sample data, i.e. hard sample data, and uses the mined first sample data to You can train the model, and by improving the proportion of hard samples in the training sample, you can reduce redundant samples and improve model training efficiency, while making the model pay more attention to hard samples and improve the model's detection performance. Let

In step 105, based on the first sample data, the training set and test set are updated to further train the liveness detection model.

Updating the training set and the test set based on the first sample data includes extracting a part of the sample data in the first sample data, adding it to the training set and the test set, and adding it to the updated training set and the updated training set. including constructing a test set for

Embodiments of the present disclosure are used throughout the training process by extracting some sample data in each of the first or hard sample data to form an updated training set and an updated test set. The data is not a complete new data plus the original data, but a partial sample in it, and it may be a hard sample, and by improving the ratio of hard samples in the training sample, reducing redundant samples, It makes the model more focused on hard samples and improves the model's detection performance while improving the model training efficiency.

Extracting a part of each sample data from the first sample data and adding them to the training set and the test set to form an updated training set and an updated test set is preset in the first sample data. extracting the second sample data at the same sampling rate, adding the extracted second sample data to the training set to form an updated training set, adding the first sample data excluding the second sample data to the test set, configuring an updated test set; decaying updating a preset sampling rate with a preset decay rate to obtain a decaying updated sampling rate; Training a liveness detection model based on a test set, and performing the training steps iteratively, stopping training when we determine that the liveness detection model has converged to a preset accuracy, and finally trained and outputting a liveness detection model.

In an embodiment of the present disclosure, the preset sampling rate takes a value greater than 0% and less than 30%. Of course, in actual application, other settings may be made to the preset sampling rate as needed, and the present disclosure is not limited thereto. By giving reference of sampling rate values, it shows implementation of specific embodiments of the present disclosure, provides guidelines for practical applications, and facilitates selection in practical applications.

The data used in the entire training process is not the complete new data plus the original data, but some samples in it, and the data is gradually added to the training set during the training process, but the added rate decays exponentially. , the data used in the training method based on hard sample mining is much smaller than the complete full data, which greatly improves the training efficiency of the model.

As shown in FIG. 2, it is a diagram illustrating a method for training a liveness detection model according to a second embodiment of the present invention. The liveness detection model training method includes the following steps.

を設定し、減衰率εを設定し、新データ＋元データの混合データセットΦを含むモデルトレーニングの入力を与え、ハイパーパラメータの初期サンプリングレートρを設定し、モデルの初期開始トレーニングのトレーニングセットの分割に使用し、データセットΦにおいて、ハイパーパラメータの初期サンプリングレートρで定量のデータをランダムにサンプリングして初期トレーニングセットΦ_{ｔｒａｉｎ}を構成し、ハイパーパラメータの初期サンプリングレートρは０％より大きくかつ５０％より小さい値をとる。ハードサンプルレート
（外２）

を設定し、データにおけるハードサンプルの割合を仮定するために用いられ、０％より大きくかつ３０％より小さい値をとる。減衰率εを設定し、ハードサンプルレートの減衰を特徴付けるために用いられ、モデルが次第に収束するにつれて、データ中に残っているハードサンプルの割合が次第に減少し、したがって、減衰率はハードサンプルレートの減衰程度を特徴付けるために用いられる。ここでは、ハイパーパラメータの初期サンプリングレート、ハードサンプルレート、減衰率はすべて応用ニーズに応じて人為的に設定することができる。このようにして様々な異なる応用ニーズを満たすことができ、より良いモデルトレーニング効果を達成する。 In step 201, we input the data set Φ, the initial sampling rate ρ, and the hard sample rate (outer 1)

, set the decay rate ε, give an input for model training containing a mixed dataset Φ of new + original data, set the initial sampling rate ρ for the hyperparameters, set the training set for the initial start training of the model used for splitting, constructing an initial training set Φ _train by randomly sampling quantitative data at an initial hyperparameter sampling rate ρ in the dataset Φ, where the initial hyperparameter sampling rate ρ is greater than 0% and 50 %. Hard sample rate (outer 2)

and is used to assume a hard sample fraction in the data, taking values greater than 0% and less than 30%. We set the decay rate ε, which is used to characterize the decay of the hard sample rate, and as the model gradually converges, the fraction of hard samples remaining in the data gradually decreases, thus the decay rate is the hard sample rate. Used to characterize the degree of attenuation. Here, the hyperparameter initial sampling rate, hard sampling rate and decay rate can all be set artificially according to application needs. In this way, it can meet various different application needs and achieve better model training effect.

を構成し、残りのデータは初期テストセット
（外４）

を構成するために使用される。 In step 202, the data set Φ is uniformly used based on the hyperparameter initial sampling rate ρ to generate an initial training set (outer 3)

and the rest of the data is the initial test set (outer 4)

used to construct the

In step 203, train the model in Φ _train , test the model in Φ _test , and sort the samples in Φ _test according to their prediction scores.

The model is first trained for a fixed number of iterations on the training set Φ _train , then the model is tested on the initial test set, scoring the sample prediction scores for the samples in the test set, sorting the prediction scores, sorting The purpose of doing is to select a hard sample, and by setting a threshold, we can establish a hard sample that meets this predetermined threshold requirement. By setting the hard sample determination method in this way, the hard samples can be found better. Theoretically, a positive sample has a prediction score of 1, a negative sample has a prediction score of 0, and a positive sample has a prediction score much less than 1 (e.g., less than 0.5). , this sample is considered a hard sample. Similarly, if the prediction score of a negative sample is much greater than 0, then this sample is also considered a hard sample.

；
（外６）

；
（外７）

のように更新する。 At step 204, the samples in Φ _test are hard sampled, and the training and test sets are
(outside 5)

;
(Outside 6)

;
(outside 7)

update like

であり、抽出されたサンプルをトレーニングセットに戻し、それによってトレーニングセット
（外９）

とテストセット
（外１０）

を更新し、それと同時に、ハードサンプルレートを
（外１１）

のように減衰更新する。 According to the set hard sample rate, in Φ _test , based on the prediction score, extract positive samples with low prediction scores and negative samples with high prediction scores, and the sampling rate is the hard sample rate (outer 8)

, returning the extracted samples to the training set, whereby the training set (outer 9)

and test set (outside 10)

and at the same time change the hard sample rate to (outside 11)

Update decay as follows.

In step 205, it is determined whether the model has converged to a preset accuracy, and if so, the process ends and training is stopped; otherwise, the process proceeds to step 203 for execution.

Because the model uses a method called gradual hard sample mining, the extracted samples have a high probability of what the model considers to be hard samples. The percentage of hard samples in the set gradually increases. Since hard sample training is effective, the preset accuracy of the model is significantly improved.

The key point of this disclosure is the modeling process of hard sample mining into the model. Through the idea of progressive hard sample mining, improve the proportion of hard samples in training samples and reduce redundant samples in the training set, thereby Improve model training efficiency. It also improves the performance of the model, i.e., it improves the accuracy rate of model prediction, because it uses hard samples to achieve better training effect. Compared with the traditional training scheme, the training method of progressive hard sample mining can be used to greatly reduce the training cost, and the liveness detection model can be continuously and quickly iteratively optimized at a later stage.

It can be seen from the above that the present disclosure designs a method for training a liveness detection model based on hard sample mining, which realizes gradually selecting and training hard samples dynamically from end to end. It can effectively reduce the large number of redundant non-instructive samples, which can improve the training efficiency of liveness detection and further improve the performance of the detection model.

Face biometric detection is one of the basic technologies in the face-related field, and is applied to many scenes such as safety, attendance inspection, finance, and security passage. It is widely applied to many current tasks. The present disclosure is used to reduce the cost of late-stage optimization of biological models, improve optimization efficiency, and provide maximum model growth effects with increasing data. In order to improve training efficiency, the training time is reduced, the higher the efficiency, the more data can be trained better, the more data can be trained, the better the data growth, the more business It is advantageous for further dissemination of the item.

The training method designed in this disclosure can be applied to the optimization of any facial biometric deep learning neural network model, which helps improve the optimization efficiency of the model. This method improves the performance of face biometric detection model by increasing the hard sample ratio, and can be applied to the application scene of face biometric detection model, which needs to be continuously optimized regularly.

As shown in FIG. 3, it is a diagram showing the configuration of a liveness detection model training device according to a third embodiment of the present disclosure, wherein the liveness detection model training device comprises a training set and a test set for training a liveness detection model. a building module 301 for building a set, training a preset liveness detection model based on the training set to obtain a first liveness detection model, and training a first liveness detection model based on the test set; a training module 302 for generating test results; a sample acquisition module 303 for analyzing the test set based on the test results and obtaining first sample data; and based on the first sample data, and an update module 304 for updating the training set, test set, and further training the liveness detection model.

In an embodiment of the present disclosure, the construction module 301 constructing a training set and a test set for training a liveness detection model includes constructing a training set and a test set based on a mixed data set, wherein said A mixed dataset contains new data and original data.

In an embodiment of the present disclosure, the construction module 301 constructing a training set and a test set based on a mixed data set includes performing random sampling on the mixed data set and constructing the sampled data as a training set. and forming data other than the sampled data as a test set.

In an embodiment of the present disclosure, the sample acquisition module 303 analyzing the test set based on the test results and acquiring first sample data includes scoring and predicting data in the test set based on the test results. Obtaining a score, sorting the data in the test set based on the predicted score, and determining data for which the predicted score meets a predetermined threshold as first sample data.

In an embodiment of the present disclosure, updating the training set and the test set by the sample acquisition module 303 based on the first sample data means extracting a part of the sample data in the first sample data to obtain a training set and a test set, respectively. In addition to the set, it includes constructing an updated training set and an updated test set.

In an embodiment of the present disclosure, the sample acquisition module 303 respectively extracts a portion of sample data from the first sample data and adds them to the training set and the test set to form an updated training set and an updated test set. That is, in the first sample data, extract the second sample data at a preset sampling rate, add the extracted second sample data to the training set, constitute the updated training set, and the second sample data adding the first sample data excluding the data to the test set to form an updated test set; and decay updating the preset sampling rate with a preset decay rate to obtain the sampling rate after the decay update. training a liveness detection model based on the updated training set and the updated test set; and performing the training step iteratively to determine that the liveness detection model has converged to a preset accuracy. Then stop training and output the final trained liveness detection model.

In an embodiment of the present disclosure, performing random sampling on the mixed data set by the construction module 301 includes performing random sampling on the mixed data set based on an initial sampling rate of a preset hyperparameter. .

In an embodiment of the present disclosure, the construction module 301 performs random sampling on the mixed data set based on a preset initial sampling rate of hyperparameters, wherein the initial sampling rate of hyperparameters is less than 0%. Including taking values that are greater and less than 50%.

In an embodiment of the present disclosure, the sample acquisition module 303 is used to decay update a preset sample rate with a preset decay rate, which takes a value greater than 0% and less than 30%.

In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of personal information of users comply with relevant laws and regulations and do not violate public order and morals.

According to embodiments of the disclosure, the disclosure further provides an electronic device, a non-transitory computer-readable storage medium having computer instructions stored thereon, and a computer program product.

FIG. 4 is a block diagram illustrating an exemplary electronic device 400 with which embodiments of the present disclosure may be implemented. Electronic equipment refers to various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also refer to various forms of mobile devices such as personal digital processors, cellular phones, smart phones, wearable devices and other similar computing devices. The components, their connections and their functionality shown herein are merely exemplary and do not limit the implementation of the disclosure as described and/or claimed herein.

As shown in FIG. 4, the apparatus 400 includes a computing unit 401 which can be read by a computer program stored in read-only memory (ROM) 402 or loaded from storage unit 408 into random access memory (RAM) 403 . , can perform various suitable operations and processes. RAM 403 may also store various programs and data necessary to operate instrument 400 . Computing unit 401 , ROM 402 and RAM 403 are connected together by bus 404 . Input/output (I/O) interface 405 is also connected to bus 404 .

A plurality of components in the device 400 are connected to an I/O interface 405, and a plurality of components include an input unit 406 such as a keyboard, a mouse, etc., an output unit 407 such as various types of displays, speakers, etc., a storage unit such as a magnetic disk, an optical disk, etc. It includes a unit 408 and a communication unit 409 such as a network card, modem, wireless communication transceiver. Communication unit 409 enables device 400 to exchange information/data with other devices via computer networks, such as the Internet, and/or various telecommunications networks.

Computing unit 401 may be various general purpose and/or special purpose processing components having processing and computing capabilities. Examples of computing units 401 include central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, computing units that run various machine learning model algorithms, digital signal processors ( DSP), and any suitable processor, controller, microcontroller, or the like. The computing unit 401 executes each of the methods and processes described above, such as the liveness detection module training method. For example, in some embodiments the liveness detection module training method may be implemented as a computer software program, tangibly contained in a machine-readable medium, eg, storage unit 408 . In some embodiments, some or all of the computer programs may be loaded and/or installed on device 400 via ROM 402 and/or communication unit 409 . When the computer program is loaded into RAM 403 and executed by computing unit 401, one or more steps of the liveness detection module training method described above may be performed. Alternatively, in another embodiment, computing unit 401 may be configured (eg, by firmware) to execute the liveness detection module training method in any other suitable manner.

Various embodiments of the systems and techniques described herein above are digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSP), system-on-chip (SOC), load programmable logic device (CPLD), software hardware, firmware, software, and/or combinations thereof. These various embodiments may be embodied in one or more computer programs, which may be executed and/or interpreted by a programmable system including at least one programmable processor; The programmable processor may be a dedicated or general purpose programmable processor, receives data and instructions from the storage system, at least one input device, and at least one output device, and transmits data and instructions to the storage system, the at least one input device. , may be sent to the at least one output device.

Program code implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general purpose computer, special purpose computer or other programmable data processing apparatus such that the program code, when executed by the processor or controller, is represented by the flowcharts and/or block diagrams as defined in the flowcharts and/or block diagrams. Perform a function/operation. The program code may be fully machine executed, partially machine executed, partially machine executed and partially remote machine executed as an independent software package, or fully remote machine executed. Or it may run on a server.

In the context of the present disclosure, a machine-readable medium may be a tangible medium that contains or stores a program that is used with or coupled to an instruction execution system, apparatus or device. You may A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specific examples of machine-readable storage media are electrical connections via one or more leads, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable readout including dedicated memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.

A computer may implement the systems and techniques described herein to provide interaction with a user, and the computer may include a display device (e.g., a CRT (Cathode Ray Tube)) for displaying information to the user. or a LCD (Liquid Crystal Display) monitor) and a keyboard and pointing device (eg, mouse or trackball) through which a user may input into the computer. Other types of devices may also provide user interaction. For example, the feedback provided to the user can be any form of sensory feedback (e.g., visual, auditory, or tactile feedback), and any form of feedback from the user (including audible, audio, or tactile input). may receive input from

The systems and techniques described herein may be computing systems including backstage components (e.g., as data servers), computing systems including middleware components (e.g., application servers), or computing systems including front-end components (e.g., graphical a user computer having a user interface or web browser, through which a user can interact with an embodiment of those systems or techniques), or their backstage components, middleware components, or It may be implemented in a computing system consisting of any combination of front end components. The components of the system can be interconnected by any form or medium of digital data communication (eg, a communication network). Examples of communication networks include, for example, local networks (LAN), wide area networks (WAN), and the Internet.

The computer system can include client sites and servers. A client and server are generally remote from each other and typically interact through a communication network. A client-server relationship is created by running computer programs on corresponding computers that have a client-server relationship to each other. The server may be a cloud server, a server in a distributed system, or a server in combination with a blockchain.

It should be understood that steps may be reordered, increased or deleted using the various forms of flow described above. For example, each step described in this disclosure may be executed in parallel, sequentially, or in a different order. The text is not limited to this, as long as it achieves the desired result.

The specific embodiments described above are not intended to limit the scope of the claims of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made based on design requirements and other factors. Any modification, equivalent replacement, improvement, etc. without departing from the spirit and principle of this disclosure shall fall within the protection scope of this disclosure.

Claims (21)

A method of training a liveness detection model, comprising:
constructing a training set and a test set for training a liveness detection model;
training a preset liveness detection model based on the training set to obtain a first liveness detection model;
training a first liveness detection model based on the test set to generate test results;
analyzing the test set based on the test results to obtain first sample data;
A method of training a liveness detection model, comprising updating a training set, a test set, and further training the liveness detection model based on first sample data. Constructing a training set and a test set for training the liveness detection model includes constructing a training set and a test set based on a mixed data set, the mixed data set including new data and original data. The method of claim 1. Constructing a training set and a test set based on the mixed data set includes: performing random sampling on the mixed data set; configuring the sampled data as a training set; as a test set. Analyzing the test set based on the test results to obtain first sample data includes:
Grading the data in the test set based on the test result to obtain a prediction score, sorting the data in the test set based on the prediction score, and determining the data whose prediction score satisfies a predetermined threshold as first sample data. 2. The method of claim 1, comprising: Updating the training set and the test set based on the first sample data includes extracting a part of the sample data in the first sample data, adding it to the training set and the test set, and adding it to the updated training set and the updated training set. 4. The method of claim 3, comprising constructing a test set for . Extracting a part of each sample data from the first sample data and adding them to the training set and the test set to form an updated training set and an updated test set is preset in the first sample data. extracting the second sample data at the same sampling rate, adding the extracted second sample data to the training set to form an updated training set, adding the first sample data excluding the second sample data to the test set, configuring an updated test set; decaying updating a preset sampling rate with a preset decay rate to obtain a decaying updated sampling rate; Training a liveness detection model based on a test set, performing the training steps iteratively, and when we determine that the liveness detection model has converged to a preset accuracy, stopping training and training the liveness detection model 6. The method of claim 5, comprising outputting a . 4. The method of claim 3, wherein random sampling in the mixed data set comprises random sampling in the mixed data set based on an initial sampling rate of preset hyperparameters. 8. The method of claim 7, wherein the hyperparameter initial sampling rate takes a value greater than 0% and less than 50%. 7. The method of claim 6, wherein the preset sampling rate takes a value greater than 0% and less than 30%. A liveness detection model training device comprising:
a building module for building a training set and a test set for training a liveness detection model;
A training module for training a preset liveness detection model based on the training set to obtain a first liveness detection model, training the first liveness detection model based on the test set, and generating test results. When,
a sample acquisition module for analyzing the test set based on the test results to acquire first sample data;
An apparatus for training a liveness detection model, comprising: an update module for updating a training set, a test set, and further training the liveness detection model based on first sample data. Constructing a training set and a test set for the construction module to train a liveness detection model includes constructing a training set and a test set based on a mixed data set, the mixed data set being new data and original data. 11. The device of claim 10, comprising data. The construction module constructing a training set and a test set based on a mixed data set includes: performing random sampling on the mixed data set; forming sampled data as a training set; 12. The apparatus of claim 11, comprising constructing the data other than as the test set. The sample acquisition module analyzing the test set based on the test results to acquire first sample data includes:
Grading the data in the test set based on the test result to obtain a prediction score, sorting the data in the test set based on the prediction score, and determining the data whose prediction score satisfies a predetermined threshold as first sample data. 12. The apparatus of claim 11, comprising: The sample acquisition module updating the training set and the test set based on the first sample data extracts a part of the sample data from the first sample data, adds it to the training set and the test set, and adds it to the training set after updating. 13. The apparatus of claim 12, comprising constructing a set and an updated test set. The sample acquisition module extracts some sample data from the first sample data and adds them to a training set and a test set to form an updated training set and an updated test set, wherein the first sample data In, extracting the second sample data at a preset sampling rate, adding the extracted second sample data to the training set, forming an updated training set, and removing the first sample data except the second sample data In addition to the test set, constructing an updated test set; decaying updating the preset sampling rate with a preset decay rate to obtain the post-decay updating sampling rate; Training a liveness detection model based on a set and an updated test set, and performing the training steps iteratively, stopping training when we determine that the liveness detection model has converged to a preset accuracy, and a final and outputting a dynamically trained liveness detection model. 13. The apparatus of claim 12, wherein the construction module performing random sampling on the mixed data set comprises performing random sampling on the mixed data set based on an initial sampling rate of preset hyperparameters. The construction module performing random sampling on the mixed data set based on a preset hyperparameter initial sampling rate determines that the hyperparameter initial sampling rate is greater than 0% and less than 50%. 17. The apparatus of claim 16, comprising taking. 16. The apparatus of claim 15, wherein the sample acquisition module is used to decay update a preset sample rate greater than 0% and less than 30% with a preset decay rate. an electronic device,
at least one processor; and memory communicatively coupled to the at least one processor;
The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the method of claim 1. machine. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method of claim 1. A computer program product comprising a computer program that, when executed by a processor, implements the method of claim 1.

Images