A liveness detection method and system

Abstract

This specification provides a liveness detection method and system, relating to the field of biometric technology. The liveness detection method and system provided in this specification perform liveness detection by combining user click operations with facial recognition. This eliminates the need for users to perform specific facial movements to complete additional information entry in any situation (such as in public), reducing the embarrassment users experience during liveness detection. Furthermore, by using screen click data generated by user click operations as an auxiliary means and integrating it into the facial recognition liveness detection process, the security of liveness detection can be improved through the detection and recognition of screen click data.

Description

Technical Field

[0001] This specification relates to the field of biometric technology, and in particular to a liveness detection method and system. Background Technology

[0002] Liveness detection is one of the most common security risks faced by facial recognition systems. With the widespread adoption of facial recognition applications in recent years, high-precision liveness detection has become increasingly important. However, high-precision liveness detection methods still face numerous problems and challenges.

[0003] Therefore, there is a need to provide a more efficient method and system for liveness detection. Summary of the Invention

[0004] The main purpose of this specification is to provide a method and system for live detection.

[0005] In a first aspect, this specification provides a liveness detection method, comprising: obtaining a first facial image of a target object, and inputting the first facial image into a liveness detection model to obtain click point indication information, the click point indication information being used to guide the target object to perform a screen click operation; obtaining screen click data of a screen click object and a second facial image of the screen click object, wherein the screen click object includes the target object, the screen click data reflects the habitual characteristics of the screen click object during the screen click operation according to the click point indication information, and the second facial image reflects the identity characteristics of the screen click object during the screen click operation; determining a liveness detection result of the target object based on the second facial image and the screen click data; and outputting the liveness detection result.

[0006] In some embodiments, the liveness detection model includes a click point distribution generation network; and the step of inputting the first facial image into the liveness detection model to obtain the click point indication information includes: using the click point distribution generation network to predict the first facial image to obtain a click point distribution prediction result for the target object, the click point distribution prediction result reflecting the click probability of the target object on each click area on the screen, and determining the click point indication information based on the click point distribution prediction result.

[0007] In some embodiments, the click point distribution generation network includes a facial feature extraction subnetwork and a distribution map prediction subnetwork; and the step of using the click point distribution generation network to predict the first facial image to obtain the click point distribution prediction result of the target object includes: using the facial feature extraction subnetwork to extract features from the first facial image to obtain a facial feature map, and using the distribution map prediction subnetwork to predict the facial feature map to obtain the click point distribution prediction result.

[0008] In some embodiments, the click point distribution generation network further includes a gaze pose prediction subnetwork; and the step of using the distribution map prediction subnetwork to predict the facial feature map to obtain the click point distribution prediction result includes: using the gaze pose prediction subnetwork to extract features from the facial feature map to obtain a gaze pose prediction result, and using the distribution map prediction subnetwork to predict the facial feature map and the gaze pose prediction result to obtain the click point distribution prediction result.

[0009] In some embodiments, the screen click data includes at least one of the following: click location information, click time interval, click duration, and click force when the screen click object performs the screen click operation.

[0010] In some embodiments, the liveness detection model further includes a click pattern feature map extraction network and a multimodal liveness detection network; and the step of determining the liveness detection result of the target object based on the second facial image and the screen click data includes: using the click pattern feature map extraction network to extract features from the screen click data to obtain a click pattern feature map of the screen click object, and using the multimodal liveness detection network to predict the second facial image and the click pattern feature map to obtain the liveness detection result of the target object.

[0011] In some embodiments, the multimodal liveness detection network includes a facial feature encoding subnetwork, a click feature encoding subnetwork, and a fusion decision subnetwork; and the step of using the multimodal liveness detection network to predict the second facial image and the click pattern feature map to obtain the liveness detection result of the target object includes: using the facial feature encoding subnetwork to extract features from the second facial image to obtain a facial decision feature map, using the click feature encoding subnetwork to extract features from the click pattern feature map to obtain a click liveness feature map, and using the fusion decision subnetwork to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object.

[0012] In some embodiments, the step of using the fusion decision sub-network to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object includes: fusing the facial decision feature map and the click liveness feature map to obtain a fused feature map; and using the fusion decision sub-network to predict the fused feature map to obtain the liveness detection result of the target object.

[0013] Secondly, this specification provides a liveness detection system, comprising: at least one storage medium including at least one instruction set for implementing and analyzing a liveness detection method; and at least one processor communicatively connected to the at least one storage medium, wherein, when the system is running, the at least one processor reads the at least one instruction set and executes the method as described above according to the instructions of the at least one instruction set.

[0014] Thirdly, this specification provides a training method for a liveness detection model, comprising: obtaining a first image sample of a training object, a screen click sample of a screen click object, and a second image sample of the screen click object, wherein the screen click object includes the training object; inputting the first image sample, the screen click sample, and the second image sample into a preset liveness detection model to obtain a click point distribution prediction result, a click prediction result, and a liveness prediction result; and training the preset liveness detection model based on the click point distribution prediction result, the click prediction result, and the liveness prediction result to obtain a trained liveness detection model.

[0015] In some embodiments, the preset liveness detection model includes a preset click point distribution generation network, a preset click pattern feature map extraction network, and a preset multimodal liveness detection network; and the step of inputting the first image sample, the screen click sample, and the second image sample into the preset liveness detection model to obtain a click point distribution prediction result, a click pattern feature map, and a facial decision feature map includes: inputting the first image sample into the preset click point distribution generation network to obtain the click point distribution prediction result; inputting the screen click sample into the preset click pattern feature map extraction network to obtain the click pattern feature map and the click prediction result; and inputting the click pattern feature map and the second image sample into the preset multimodal liveness detection network to obtain the liveness prediction result.

[0016] In some embodiments, the preset click pattern feature map extraction network includes a preset click feature extraction subnetwork, a preset identity classification subnetwork, and a preset click point distribution regression subnetwork; and the step of inputting the screen click sample into the preset click pattern feature map extraction network to obtain the click pattern feature map and the click prediction result includes: using the preset click feature extraction subnetwork to extract features from the screen click sample to obtain the click pattern feature map, using the preset identity classification subnetwork to predict the click pattern feature map to obtain an identity prediction result, and using the preset click point distribution regression subnetwork to predict the click pattern feature map to obtain a click point distribution regression prediction result, wherein the click prediction result includes the identity prediction result and the click point distribution regression prediction result.

[0017] In some embodiments, the preset multimodal liveness detection network includes a preset facial feature encoding subnetwork, a preset click feature encoding subnetwork, and a preset liveness decision subnetwork; and the step of inputting the click pattern feature map and the second image sample into the preset multimodal liveness detection network to obtain the liveness prediction result includes: using the preset facial feature encoding subnetwork to extract features from the second image sample to obtain a facial decision feature map, using the preset click feature encoding subnetwork to extract features from the click pattern feature map to obtain a click liveness feature map, and using the preset liveness decision subnetwork to predict the facial decision feature map and the click liveness feature map to obtain the liveness prediction result.

[0018] In some embodiments, the preset liveness decision subnetwork includes a preset facial feature decision subnetwork, a preset click liveness feature decision subnetwork, and a preset fusion decision subnetwork; and the prediction of the facial decision feature map and the click liveness feature map based on the preset liveness decision subnetwork to obtain the liveness prediction result includes: using the preset facial feature decision subnetwork to predict the facial decision feature map to obtain a first liveness prediction result, using the preset click liveness feature decision subnetwork to predict the click liveness feature map to obtain a second liveness prediction result, and using the preset fusion decision subnetwork to predict the facial decision feature map and the click liveness feature map to obtain a third liveness prediction result, wherein the liveness prediction result includes the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result.

[0019] In some embodiments, training the preset liveness detection model based on the click point distribution prediction result, the click prediction result, and the liveness prediction result includes: training the preset click point distribution generation network based on the click point distribution prediction result to obtain a trained click point distribution generation network; training the preset click pattern feature map extraction network based on the click prediction result to obtain a trained click pattern feature map extraction network; and training the preset multimodal liveness detection network based on the liveness prediction result to obtain a trained multimodal liveness detection network.

[0020] In some embodiments, training the preset click point distribution generation network based on the click point distribution prediction results to obtain the trained click point distribution generation network includes: obtaining the click point distribution annotation results of the training object; comparing the click point distribution prediction results with the click point distribution annotation results to obtain first click point distribution loss information; and converging the preset click point distribution generation network based on the first click point distribution loss information to obtain the trained click point distribution generation network.

[0021] In some embodiments, training the preset click pattern feature map extraction network based on the click prediction result to obtain the trained click pattern feature map extraction network includes: obtaining the identity labeling result of the training object, and comparing the identity prediction result with the identity labeling result to obtain identity classification loss information; obtaining the click point distribution labeling result of the training object, and comparing the click point distribution regression prediction result with the click point distribution labeling result to obtain second click point distribution loss information; and converging the preset click pattern feature map extraction network based on the identity classification loss information and the second click point distribution loss information to obtain the trained click pattern feature map extraction network.

[0022] In some embodiments, training the preset multimodal liveness detection network based on the liveness prediction results to obtain the trained multimodal liveness detection network includes: determining liveness classification loss information of the multimodal liveness detection network based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result; comparing the first liveness prediction result and the third liveness prediction result to obtain first consistency prediction loss information; comparing the second liveness prediction result and the third liveness prediction result to obtain second consistency prediction loss information; and converging the preset multimodal liveness detection network based on the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information to obtain the trained multimodal liveness detection network.

[0023] In some embodiments, the liveness classification loss information includes first liveness classification loss information, second liveness classification loss information, and third liveness classification loss information; and determining the liveness classification loss information of the multimodal liveness detection network based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction includes: obtaining a first liveness labeling result of the training object based on facial features, and comparing the first liveness prediction result with the first liveness labeling result to obtain the first liveness classification loss information; obtaining a second liveness labeling result of the training object based on click point operations, and comparing the second liveness prediction result with the second liveness labeling result to obtain the second liveness classification loss information; and obtaining a third liveness labeling result of the training object based on the facial features and the click point operations, and comparing the third liveness prediction result with the third liveness labeling result to obtain the third liveness classification loss information.

[0024] Fourthly, this specification provides a training system for a liveness detection model, comprising: at least one storage medium including at least one instruction set for implementing and analyzing a training method for the liveness detection model; and at least one processor communicatively connected to the at least one storage medium, wherein, when the system is running, the at least one processor reads the at least one instruction set and executes the method as described above according to the instructions of the at least one instruction set.

[0025] As can be seen from the above technical solutions, the liveness detection method and system provided in this specification perform liveness detection by combining user click operations with facial recognition. This eliminates the need for users to perform specific facial movements to complete additional information entry in any situation (such as public places), reducing the embarrassment users experience during liveness detection. Furthermore, by using screen click data generated by user click operations as an auxiliary means and integrating it into the liveness detection process of facial recognition, the security of liveness detection can be improved through the detection and recognition of screen click data.

[0026] Other functions of the liveness detection methods and systems provided in this specification will be partially listed in the following description. The figures and examples described below will be readily apparent to those skilled in the art. The inventive aspects of the liveness detection methods and systems provided in this specification can be fully understood through practice or use of the methods, apparatus, and combinations described in the detailed examples below. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A schematic diagram illustrating an application scenario of a liveness detection method according to some embodiments of this specification is shown.

[0029] Figure 2 A schematic diagram of the structure of a computing device provided according to some embodiments of this specification is shown;

[0030] Figure 3 A flowchart of a liveness detection method according to some embodiments of this specification is shown;

[0031] Figure 4 This diagram illustrates a display method for click point indication information in a liveness detection method according to some embodiments of this specification;

[0032] Figure 5 A schematic diagram of a liveness detection model provided according to some embodiments of this specification is shown;

[0033] Figure 6 A schematic diagram of the data flow of a click point distribution generation network provided according to some embodiments of this specification is shown.

[0034] Figure 7 A schematic diagram of the data flow of a click point distribution generation network provided according to some embodiments of this specification is shown.

[0035] Figure 8 A schematic diagram of the data flow of a click pattern feature map extraction network provided according to some embodiments of this specification is shown.

[0036] Figure 9 This diagram illustrates the data flow of a multimodal liveness detection network provided according to some embodiments of this specification.

[0037] Figure 10 A flowchart illustrating a method for training a liveness detection model according to some embodiments of this specification is shown.

[0038] Figure 11 A flowchart illustrating a method for training a liveness detection model according to some embodiments of this specification is shown.

[0039] Figure 12A schematic diagram of the data flow of a preset click pattern feature map extraction network provided according to some embodiments of this specification is shown.

[0040] Figure 13 This diagram illustrates the data flow of a preset multimodal liveness detection network provided according to some embodiments of this specification; and

[0041] Figure 14 A flowchart illustrating a method for training a liveness detection model according to some embodiments of this specification is shown. Detailed Implementation

[0042] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0043] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0044] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0045] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0046] The collection, use, and processing of data involved in the liveness detection described in this instruction manual have been authorized and agreed to by the user and comply with relevant national laws and regulations.

[0047] Figure 1 A schematic diagram illustrating an application scenario of a system 100 for a liveness detection method according to some embodiments of this specification is shown. The system 100 for the liveness detection method (hereinafter referred to as system 100) may include a client 110, an integrated development platform server 130, and a database 140.

[0048] Integrated Development Platform (IDE) 120 is mounted on IDE server 130. IDE 120, also known as Integrated Development Environment (IDE), is an application that provides a program development environment, generally including tools such as a code editor, compiler, debugger, and graphical user interface. Developers can write program code (i.e., program development) on IDE 120 through client 110. IDE server 130 (hereinafter referred to as server 130) can be a computing device on IDE 120 specifically used to process the liveness detection method. Server 130 can store data or instructions for executing the liveness detection method described in this specification, and can execute or be used to execute said data and / or instructions. Server 130 may include hardware devices with data information processing capabilities and the necessary programs required to drive the hardware devices. Of course, server 130 may also be only a hardware device with data processing capabilities, or only a program running on the hardware device. In some embodiments, server 130 may also be deployed as a plug-in on client 110.

[0049] Client 110 may include at least one mobile terminal device. For example, the number of clients 110 may be one, two, three, four, five, or other numbers. The at least one mobile terminal device may include a mobile terminal device with liveness detection functionality. The at least one mobile terminal device may be a mobile terminal device that simultaneously has image acquisition and touch display functionality. For example, the at least one mobile terminal device may be a computer, mobile phone, tablet, or other device with a camera and a touch display screen. The mobile terminal device can display click point indication information on the touch display screen, allowing the user to perform screen click operations based on the click point indication information. Simultaneously, the mobile terminal device can acquire screen click data corresponding to the user's screen click operation. The mobile terminal device can acquire the user's image information through a camera. For example, the mobile terminal device can acquire the user's facial image information or other identifiable body part image information through a camera. Client 110 may also include an application on at least one mobile terminal device.

[0050] Database 140 may store data and / or instructions. In some embodiments, database 140 may store data and / or instructions executed by server 130 or used to execute the liveness detection methods described herein. Client 110 and server 130 may have access to database 140, and client 110 and server 130 may access data or instructions stored in database 140 via a network. In some embodiments, database 140 may be directly connected to client 110 and server 130. In some embodiments, database 140 may be part of server 130. In some embodiments, database 140 may include mass storage, removable storage, volatile read-write memory, read-only memory (ROM), or similar content, or any combination thereof. Exemplary mass storage may include non-transitory storage media such as disks, optical discs, and solid-state drives. Exemplary removable storage may include flash drives, floppy disks, optical discs, memory cards, zip disks, magnetic tapes, etc. Typical volatile read-write memory may include random access memory (RAM). Example RAMs may include dynamic RAM (DRAM), dual date rate synchronous dynamic RAM (DDRSDRAM), static RAM (SRAM), thyristor RAM (T-RAM), and zero-capacitance RAM (Z-RAM), etc. Exemplary ROMs may include mask ROM (MROM), programmable ROM (PROM), virtual programmable ROM (PEROM), electronically programmable ROM (EEPROM), optical disc (CD-ROM), and digital multifunction disk ROM, etc.

[0051] It should be understood that Figure 1The number of clients 110 and servers 130 shown is merely illustrative. Depending on implementation needs, there can be any number of clients 110 and servers 130.

[0052] It should be noted that the liveness detection method can be executed entirely on the client 110, entirely on the server 130, or partially on the client 110 and partially on the server 130.

[0053] For ease of description, the following descriptions will use the execution of the liveness detection method on server 130 as an example to describe the technical solutions involved in this specification.

[0054] Figure 2 This is a schematic diagram of the structure of a computing device 200 provided according to some embodiments of this specification. The computing device 200 can be a general-purpose computer or a special-purpose computer. For example, the computing device 200 can be a server, a personal computer, a portable computer (such as a laptop computer, tablet computer, etc.), or other electronic devices with computing capabilities. Of course, the computing device can be... Figure 1 The server 130 can also be a terminal device used by multiple developers to develop programs on an integrated development platform.

[0055] like Figure 2 As shown, the computing device 200 may include a COM port 250, which can be connected to or from a network to facilitate data communication. The computing device 200 may also include a processor 220 in the form of one or more processors, such as a central processing unit (CPU), for executing program instructions. The computing device 200 may also include an internal communication bus 210 and various forms of program storage media and data storage media, such as a disk 270 (non-transitory memory) and read-only memory (ROM) 230 or random access memory (RAM) 240, etc., for storing various data files to be processed and / or transmitted. The storage media may be local to the computing device 200 or shared by the computing device 200 (e.g., Figure 1 The storage medium M shown is referenced here. The computing device 200 may also include program instructions stored in ROM 230, RAM 240, and / or other types of non-transitory storage media to be executed by processor 220. The computing device 200 may also include I / O components 260 to support data communication with other computing devices in the distributed computing system 100. The computing device 200 may also receive programming and data via network communication.

[0056] For illustrative purposes only, only one processor 220 is described in the computing device 200. However, those skilled in the art will understand that the computing device 200 in this specification may also include multiple processors. Therefore, the methods / steps / operations performed by one processor as described in this specification may also be performed jointly or separately by multiple processors. For example, in this specification, the processor of the computing device 200 may simultaneously execute step A and step B. It should be understood that step A and step B may also be performed jointly by two different processors. For example, a first processor executes step A, a second processor executes step B, or a first processor and a second processor jointly execute steps A and B.

[0057] Figure 3 A flowchart 300 of a liveness detection method according to some embodiments of this specification is shown. The following will be combined with... Figure 3 This specification describes the technical solution for the live detection method. The subject implementing the described live detection method may be... Figure 1 At least one of client 110 and server 130. Specifically, the client 110 and / or server 130 may have, for example: Figure 2 The aforementioned structure, namely, the client 110 and / or server 130, can be a device for a liveness detection method, comprising: at least one storage medium and at least one processor. The at least one storage medium includes at least one instruction set for implementing the liveness detection method. The at least one processor is communicatively connected to the at least one storage medium. When the system is running, the at least one processor can read the at least one instruction set and execute instructions according to the at least one instruction set. Figure 3 The method 300. For illustrative purposes only, this specification will describe method 300 using server 130 as an example. Method 300 may include:

[0058] S310, Obtain a first facial image of the target object, and input the first facial image into a liveness detection model to obtain click point indication information, which is used to guide the target object to perform a screen click operation.

[0059] In this specification, the target object can be understood as the user currently performing liveness detection. Client 110 can acquire a first facial image of the target object and send it to server 130. After receiving the first facial image sent by client 110, server 130 can input the first facial image into the liveness detection model and output click point indication information using the liveness detection model. In this specification, server 130 can output the click point indication information to client 110 for display on client 110's screen. The target object can perform screen click operations guided by the click point indication information displayed on client 110's screen.

[0060] During liveness detection, click point indication information is displayed on the client 110 screen. Specifically, the client 110 screen can be divided into two or more independent click areas, each corresponding to a click point. Click point indication information can be understood as the pattern of displaying or highlighting the independent click areas (click points) on the client 110 screen. This information may include the number of times the click point is displayed, the time interval between two consecutive click point displays, and the display position of each click point. For different target objects, the server 130 can output different click point indication information. At different times or within different time periods, one of the independent click areas (click points) is displayed or highlighted on the client 110 screen. The difference between displaying and highlighting lies in the different ways the click points are displayed on the client 110 screen.

[0061] At different times or within different time periods, a single clickable area (click point) is displayed on the client 110 screen. This can be understood as follows: during liveness detection, all click points on the client 110 screen are hidden. When the client 110 screen does not display click point indicators, the target object cannot perceive the distribution and quantity of all click points through the client 110 screen. When the client 110 screen begins to display click point indicators, it displays only one click point at a time at different times or within different time periods to guide the target object to click on the displayed click point. Accordingly, the target object can perceive the position of the displayed click point and thus perform screen click operations according to the position of the displayed click point.

[0062] At different times or within different time periods, a single clickable area (click point) is highlighted on the client 110 screen. This can be understood as follows: during the liveness detection process, before the client 110 screen displays the click point indicator information, it displays all the click points it contains, and the display status (such as brightness and color) of each click point is the same. At this time, the target object can perceive the distribution and quantity of all click points through the client 110 screen. When the client 110 screen begins to display the click point indicator information, at different times or within different time periods, the client 110 screen highlights one click point at a time, making the highlighted click point clearly distinguishable from the other click points. The highlighting method can include highlighting, flashing, or other methods that clearly distinguish the highlighted click point from the other click points. At this time, the target object can perceive the position of the highlighted click point and thus perform screen click operations according to the position of the highlighted click point.

[0063] In this manual, the number of clickable areas on the client 110 screen can be set as needed. For example, the client 110 screen can display two independent clickable areas (clickable areas), and correspondingly, two clickable areas can be shown on the client 110 screen. The client 110 screen can also display three, four, five, or more independent clickable areas (clickable areas), and correspondingly, a corresponding number of clickable areas can be shown on the client 110 screen. In this manual, the independent clickable areas (clickable areas) on the client 110 screen can be set adjacent to each other or spaced apart. The area of ​​each independent clickable area on the client 110 screen can be the same or different.

[0064] In this specification, the number of times the click point is displayed in the click point indication information can be the same as or different from the number of click points on the client 110 screen. Furthermore, the display position of the click point can be the same or different at different times or within different time periods.

[0065] In this specification, the time interval between two consecutive displayed click points can be determined based on the time required for a user to perform a normal screen click operation. Generally, the time interval between two consecutive displayed click points is greater than or equal to the time required for a user to perform a normal screen click operation, allowing the target object ample time to complete the screen click operation; however, considering the efficiency of liveness detection, the time interval between two consecutive displayed click points cannot be too long. For example, the time interval between two consecutive displayed click points can be 0.1 seconds, 0.2 seconds, 0.3 seconds, 0.4 seconds, 0.5 seconds, 0.6 seconds, 0.7 seconds, 0.8 seconds, 0.9 seconds, 1 second, 2 seconds, 3 seconds, or other time intervals.

[0066] Figure 4This diagram illustrates a display method for click point indication information in a liveness detection method according to some embodiments of this specification. The example shown is nine independent click areas (click points) displayed on the screen of client 110. Figure 4 As shown, during the liveness detection process, nine independent click points are displayed on the client 110 screen, denoted as D1, D2, D3, D4, D5, D6, D7, D8, and D9. For a specific target object, the server 130 obtains the first facial image of the target object and, based on the liveness detection model, determines the click point indication information corresponding to the first facial image. The server then outputs the click point indication information to the client 110 for display. The click point indication information can include the number of times the click point is displayed, which can be 5 times, with a time interval of 0.5 seconds between each adjacent display. The display position of each click point can be as follows: starting from the moment the first click point appears, the first click point appears at position D2; at 0.5 seconds, the second click point appears at position D1; at 1 second, the second click point appears at position D5; at 1.5 seconds, the second click point appears at position D9; and at 2.0 seconds, the second click point appears at position D1. Accordingly, the following display occurs on the screen of client 110: at second 0, click point D2 is displayed or highlighted on the screen of client 110; at second 0.5, click point D1 is displayed or highlighted on the screen of client 110; at second 1, click point D5 is displayed or highlighted on the screen of client 110; at second 1.5, click point D9 is displayed or highlighted on the screen of client 110; and at second 2.0, click point D1 is displayed or highlighted on the screen of client 110.

[0067] It should be understood that Figure 4 The given display method for click point indication information is merely one example of the display methods used in this specification, and it does not limit the display methods of click point indication information in this specification. Other methods can also be used to display click point indication information in this specification, which will not be elaborated upon here.

[0068] Figure 5 A schematic diagram of a liveness detection model 400 provided according to some embodiments of this specification is shown.

[0069] like Figure 5As shown in this specification, the liveness detection model 400 may include a click point distribution generation network 410. The step of inputting the first facial image into the liveness detection model to obtain the click point indication information may include: using the click point distribution generation network 410 to predict the first facial image, obtaining a click point distribution prediction result for the target object, the click point distribution prediction result reflecting the click probability of the target object on various click areas on the screen, and determining the click point indication information based on the click point distribution prediction result.

[0070] In this specification, server 130 can use the click point distribution generation network 410 in the liveness detection model 400 to extract features from the first facial image of the target object, and predict the click tendency of the target object when performing screen click operations without external guidance based on the extracted feature information. Click tendency can be represented as the probability of the target object clicking on each click area (click point) on the screen when performing screen click operations. In this specification, the click point distribution prediction result is used to represent the click probability of the target object on each click area on the screen predicted by the click point distribution generation network.

[0071] Because facial features can reflect a person's temperament and behavioral habits to some extent. Temperamental characteristics can include kindness, irritability, introversion, liveliness, and gentleness. Behavioral habits can include subconscious behaviors, such as how a person uses electronic devices or how they learn new things. Taking the common practice of setting a password to unlock an electronic device as an example, when the screen displays the same number and arrangement of clickable areas, each representing a number or symbol, different people will set different passwords, and their screen-clicking habits will also differ.

[0072] Furthermore, human facial features vary from person to person; different users possess different facial characteristics. Therefore, by utilizing the inherent connection between facial features and behavioral habits, it is possible to distinguish the behavioral habits of different individuals through facial features.

[0073] Figure 6 A schematic diagram of the data flow of a click point distribution generation network 410 provided according to some embodiments of this specification is shown.

[0074] like Figure 6As shown in this specification, the click point distribution generation network 410 may include a facial feature extraction subnetwork and a distribution map prediction subnetwork. The step of using the click point distribution generation network 410 to predict the first facial image and obtain the click point distribution prediction result of the target object may include: using the facial feature extraction subnetwork to extract features from the first facial image to obtain a facial feature map, and using the distribution map prediction subnetwork to predict the facial feature map to obtain the click point distribution prediction result.

[0075] In this specification, server 130 extracts facial features from a first facial image of the target object and predicts the corresponding click point distribution based on the extracted facial feature map. The click point distribution prediction results exhibit individual differences, and the click point indication information determined by these predictions also varies. On one hand, the click point indication information predicted by server 130 closely aligns with the target object's screen clicking habits, reflecting a human-centered approach to liveness detection and enhancing the user experience. On the other hand, server 130 can also identify liveness attacks to a certain extent by providing individually differentiated click point indication information, ensuring that the click point indication information displayed on the client 110 screen differs in each liveness detection process.

[0076] Figure 7 A schematic diagram of the data flow of a click point distribution generation network 410 provided according to some embodiments of this specification is shown.

[0077] like Figure 7 As shown in this specification, the click point distribution generation network 410 may further include a gaze pose prediction sub-network. The step of using the distribution map prediction sub-network to predict the facial feature map and obtain the click point distribution prediction result may include: using the gaze pose prediction sub-network to extract features from the facial feature map and obtain a gaze pose prediction result; and using the distribution map prediction sub-network to predict both the facial feature map and the gaze pose prediction result to obtain the click point distribution prediction result.

[0078] During liveness detection, different target objects may hold the client 110 in different postures, and the same target object may also hold the client 110 in different postures at different times. The target object's posture while holding the client 110 also influences their habitual tendencies when performing screen clicks. Therefore, considering the above-mentioned potential influences, in this specification, the server 130 can further extract features from the facial feature map obtained by the facial feature extraction subnetwork using a gaze posture prediction subnetwork to obtain gaze posture prediction results. In this specification, the gaze posture prediction results can reflect the gaze and posture of the target object when performing screen clicks. The gaze posture prediction results obtained by the server 130 through the gaze posture prediction subnetwork can serve as an auxiliary condition for determining the click point distribution prediction results.

[0079] Therefore, in this specification, server 130 can directly predict the click point distribution prediction result corresponding to the target object based on the facial feature map of the target object. Server 130 can also combine the facial feature map of the target object and the gaze posture prediction result to predict the click point distribution prediction result corresponding to the target object. When server 130 combines the facial feature map of the target object and the gaze posture prediction result to predict the click point distribution prediction result corresponding to the target object, the gaze posture prediction result serves as an auxiliary condition, which can enhance the individual differences of the click point distribution prediction result and thus further enhance the individual differences of the click point indication information.

[0080] In this specification, the process by which server 130 uses a click distribution generation network 410 to predict the first facial image, obtain the click distribution prediction result of the target object, and determine the click indication information based on the click distribution prediction result can be as follows. Figure 4Taking the example of a client 110 screen containing nine clickable areas and generating verification information containing six click points, the server 130 can use a click point distribution prediction network to process the first facial image and predict the target object's screen click tendency. For example, the server 130 uses the click point distribution prediction network to process the first facial image and predicts that the target object tends to click in the following ways: the probability of the target object clicking on clickable area D1 is 15%, the probability of the target object clicking on clickable area D2 is 35%, the probability of the target object clicking on clickable area D5 is 20%, the probability of the target object clicking on clickable area D6 is 30%, and the probability of the target object clicking on other clickable areas (including D3, D4, D7, D8, and D9) is 0. Accordingly, the server 130 uses a click distribution prediction network to process the first facial image and predicts the click distribution of the target object as follows: the click probability of the target object on the click areas D1, D2, D3, D4, D5, D6, D7, D8 and D9 on the screen is 15%, 35%, 0%, 0%, 20%, 30%, 0%, 0%, 0%, respectively.

[0081] It should be understood that when a target object performs a screen click operation without external guidance, it may click the same click area multiple times. For example, in the example above, server 130 predicts that the target object will click four areas (D1, D2, D5, and D6) instead of six.

[0082] In this specification, server 130 can randomly generate click point indication information based on the predicted click point distribution. Specifically, server 130 can randomly generate click point indication information based on the predicted click probability of the target object on each clickable area on the screen. Specifically, server 130 can generate click point indication information according to the click probability of each clickable area from high to low, with clickable areas with higher click probabilities displaying their corresponding click points more frequently. In this specification, the display order of the click points in the click point indication information can be arranged according to the click probability or not. Furthermore, for multiple liveness detections of the same target object, server 130 determines the same click point distribution prediction result based on the first facial image of the target object each time, but the click point indication information generated by server 130 based on the click point distribution prediction result can be the same or different.

[0083] For example, based on the above-mentioned click point distribution prediction results, when the target object undergoes its first liveness detection, the click point indication information generated by server 130 may include: displaying D2, D5, D6, D2, D1, and D5 sequentially at 1-second intervals. When the target object undergoes its second liveness detection, the click point indication information generated by server 130 may include: displaying D5, D2, D5, D2, D1, and D6 sequentially at 1-second intervals. When the target object undergoes its third liveness detection, the click point indication information generated by server 130 may include: displaying D6, D2, D5, D2, D1, and D6 sequentially at 1-second intervals. In subsequent liveness detections, the click point indication information generated by server 130 can also be randomly adjusted, which will not be elaborated upon here.

[0084] S320, obtain screen click data of the screen click object and a second facial image of the screen click object, wherein the screen click object includes the target object, the screen click data reflects the habitual characteristics of the screen click object in the process of performing the screen click operation according to the click point indication information, and the second facial image reflects the identity characteristics of the screen click object in the process of performing the screen click operation.

[0085] As previously described, the click point indication information obtained by server 130 based on the first facial image of the target object is used to guide the target object to perform screen click operations. In this specification, during the process of server 130 outputting the click point indication information to client 110 and displaying it on the screen of client 110, after each click point is displayed on the screen of client 110, the target object can perform screen click operations according to the click points displayed on the screen of client 110. For example, taking a 1-second time interval, client 110 displays D2, D5, D6, D2, D1, and D5 sequentially. The time when the first click point D2 is displayed on the client 110 screen is counted as 0 seconds. After the first click point D2 is displayed on the client 110 screen, the target object can perform the first screen click operation. In the 1st second, after the second click point D5 is displayed on the client 110 screen, the target object can perform the second screen click operation. In the 2nd second, after the third click point D6 is displayed on the client 110 screen, the target object can perform the third screen click operation. In the 3rd second, after the fourth click point D2 is displayed on the client 110 screen, the target object can perform the fourth screen click operation. In the 4th second, after the fifth click point D1 is displayed on the client 110 screen, the target object can perform the fifth screen click operation. In the 5th second, after the sixth click point D5 is displayed on the client 110 screen, the target object can perform the sixth screen click operation.

[0086] Server 130 generates click point indication information based on the first facial image of the target object collected by client 110, and outputs the click point indication information to client 110 for display to guide the target object to perform a screen click operation. In principle, the subsequent screen click operation should be performed by the target object itself. However, in reality, due to various reasons, the subsequent screen click operation may or may not be performed by the target object itself. In this specification, for ease of description, the person performing the screen click operation is collectively referred to as the screen click object, which includes the target object.

[0087] In this specification, during the screen clicking operation, the client 110 can acquire screen clicking data corresponding to the user's screen clicking operation through its own sensors and feed the screen clicking data back to the server 130 for liveness detection. The screen clicking data may include at least one of the following: click position information, click time interval, click duration, and click force. The click position information includes the click position of the screen clicking object each time it performs a screen clicking operation. Since the click point displayed on the client 110 screen is an independent click area with a certain area, the screen clicking object can be considered to have clicked the corresponding click point as long as the finger's click position falls within the area covered by the click point. In actual screen clicking operations, although the screen clicking object's click position falls within the corresponding click point, due to personal habits, the screen clicking object may prefer to click the center or edge of the click point's area. The click time interval includes the time interval from when a click point is displayed on the client 110 screen to when the screen clicking object begins its screen clicking operation. Click duration includes the duration of contact between the finger and the screen during each click. Click force includes the force applied to the screen during each click.

[0088] In this specification, screen click data reflects the habitual characteristics of a screen-clicking object during the screen-clicking operation according to the click point indication information. Different screen-clicking objects have different habits in performing screen-clicking operations, and correspondingly, the screen click data generated by different screen-clicking objects is different. In this specification, server 130 can utilize the individual differences in the screen-clicking operation habits of different screen-clicking objects and apply them to liveness detection and recognition to improve the accuracy of liveness detection and recognition.

[0089] In this specification, the second facial image can be understood as the facial image captured by the client 110 during the screen click operation of the screen click object. The second facial image may or may not be the facial image of the target object. In this specification, the second facial image can reflect the identity characteristics of the screen click object during the screen click operation. Different screen click objects will have significantly different corresponding second facial images. The server 130 can utilize the different facial features of different screen click objects for liveness detection and recognition.

[0090] S330, the liveness detection result of the target object is determined based on the second facial image and the screen click data. In this specification, the server 130 can combine the information from the second facial image and the screen click data for data processing to predict the liveness detection result of the target object.

[0091] like Figure 5 As shown in this specification, the liveness detection model 400 may further include a click pattern feature map extraction network 420 and a multimodal liveness detection network 430. Determining the liveness detection result of the target object based on the second facial image and the screen click data may include: using the click pattern feature map extraction network 420 to extract features from the screen click data to obtain a click pattern feature map of the screen click object; and using the multimodal liveness detection network 430 to predict the liveness detection result of the target object based on the second facial image and the click pattern feature map.

[0092] Figure 8 A schematic diagram of the data flow of a click pattern feature map extraction network 420 provided according to some embodiments of this specification is shown. Figure 9 A schematic diagram of the data flow of a multimodal liveness detection network 430 provided according to some embodiments of this specification is shown.

[0093] like Figure 8 , 9 As shown in this specification, server 130 can first use click pattern feature map extraction network 420 to extract features from screen click data to obtain a click pattern feature map of the screen click object. The click pattern feature map includes the click position information, click time interval, click duration, and click force corresponding to the screen click operation. Server 130 then inputs the click pattern feature map and the second facial image into multimodal liveness detection network 430, and uses multimodal liveness detection network 430 to process the second facial image and the click pattern feature map to predict the liveness detection result of the target object.

[0094] Specifically, the multimodal liveness detection network 430 may include a facial feature encoding subnetwork, a click feature encoding subnetwork, and a fusion decision subnetwork. The step of using the multimodal liveness detection network 430 to predict the liveness detection result of the target object based on the second facial image and the click pattern feature map may include: using the facial feature encoding subnetwork to extract features from the second facial image to obtain a facial decision feature map; using the click feature encoding subnetwork to extract features from the click pattern feature map to obtain a click liveness feature map; and using the fusion decision subnetwork to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object.

[0095] In this specification, the facial decision feature map contains the facial features of the screen-clicked object, and the click liveness feature map contains the clicking habit features of the screen-clicked object when performing screen-click operations. The facial decision feature map and the click liveness feature map reflect the individual differences of the screen-clicked object from different perspectives. The server 130 can integrate the feature information contained in both the facial decision feature map and the click liveness feature map to perform liveness detection.

[0096] Taking liveness detection of a target object as an example, when the target object opens the liveness detection interface through client 110, client 110 can capture the first facial image of the target object and send it to server 130. Server 130 can make predictions based on the first facial image of the target object to obtain click point indication information corresponding to the target object, and send the click point indication information to client 110. The click point indication information is displayed on the screen of client 110 to guide the target object to perform a screen click operation.

[0097] If the screen click object is the target object, and the second facial image collected by the client 110 is the facial image of the target object, then if the screen click data generated by the target object's current screen click operation meets the guidance requirements of the click point indication information (i.e., after a certain click point is displayed on the client 110 screen, the target object clicks that click point instead of other click points), and conforms to the target object's clicking habits, then the liveness detection result of the target object predicted by the server 130 is safe.

[0098] If the screen click object is the target object, and the second facial image collected by the client 110 is the facial image of the target object, then if the screen click data generated by the target object's current screen click operation meets the guidance requirements of the click point indication information but does not meet the target object's clicking habits, then the liveness detection result of the target object predicted by the server 130 is unsafe.

[0099] If the screen click object performing the screen click operation is the target object, and the second facial image collected by the client 110 is the facial image of the target object, then if the screen click data generated by the target object's current screen click operation does not conform to the guidance requirements of the click point indication information, then regardless of whether the screen click data generated by the target object's current screen click operation conforms to the target object's clicking habits, the liveness detection result of the target object predicted by the server 130 is unsafe.

[0100] If the screen click object is not the target object, the second facial image collected by the client 110 is not the facial image of the target object. Based on this, regardless of whether the screen click data generated by the current screen click operation of the screen click object conforms to the click habits of the target object, the liveness detection result of the target object predicted by the server 130 is unsafe.

[0101] In this specification, the step of using the fusion decision sub-network to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object may include: fusing the facial decision feature map and the click liveness feature map to obtain a fused feature map; and using the fusion decision sub-network to predict the fused feature map to obtain the liveness detection result of the target object.

[0102] Although different people have different clicking habits when performing screen clicks, meaning there is a correspondence between clicking habits and identity, the stability of this correspondence is slightly lower than the one-to-one correspondence between faces and identities. Therefore, in this specification, when the server 130 uses the fusion decision sub-network to predict facial decision feature maps and click point liveness feature maps, it can prioritize the facial decision feature map and use the click point liveness feature map as a supplement. Specifically, during the feature fusion process of the facial decision feature map and the click liveness feature map, the server 130 can adjust their weights, making the weight of the facial decision feature map greater than that of the click liveness feature map, so that the fused feature map contains more facial feature information.

[0103] In this specification, server 130 performs liveness detection by combining user click operations with facial recognition. This eliminates the need for users to perform specific facial movements to complete additional information entry in any situation (such as in public), reducing the awkwardness of liveness detection for users. Furthermore, server 130 uses screen click data generated by user click operations as an auxiliary means, integrating it into the liveness detection process of facial recognition. The detection and recognition of screen click data can improve the security of liveness detection.

[0104] S340, output the liveness detection result.

[0105] In this specification, after determining the liveness detection result of the target object, the server 130 can output the liveness detection result to the client 110. The target object can know whether the liveness detection was successful through the client 110. A liveness detection result of "safe" indicates that the target object has passed the liveness detection; otherwise, it has failed. If the target object fails the liveness detection, it can re-perform liveness verification through the client 110, and the server 130 can repeat the previous liveness detection method to re-perform the liveness detection.

[0106] This specification also provides a training method for a liveness detection model. Figure 10 A flowchart 300 illustrating a method for training a liveness detection model according to some embodiments of this specification is shown. The following will be combined with... Figure 10 This specification describes the technical solution for training the liveness detection model. The entity implementing this technical solution can be... Figure 1 At least one of client 110 and server 130. Specifically, the client 110 and / or server 130 may have, for example: Figure 2 The aforementioned structure, namely, the client 110 and / or server 130, can be a device for a training method of a liveness detection model, comprising: at least one storage medium and at least one processor. The at least one storage medium includes at least one instruction set for implementing the training method of the liveness detection model. The at least one processor is communicatively connected to the at least one storage medium. When the system is running, the at least one processor can read the at least one instruction set and execute instructions according to the at least one instruction set. Figure 10 The method 500. For illustrative purposes only, this specification will describe the method 500 using server 130 as an example. The method 500 may include:

[0107] S510, obtain a first image sample of the training object, a screen click sample of the screen click object, and a second image sample of the screen click object, wherein the screen click object includes the training object.

[0108] For ease of description, the liveness detection model in the training phase is referred to as the preset liveness detection model. The training object can be understood as the sampled object before the screen click operation during the training process of the preset liveness detection model. The screen click object can be understood as the sampled object during the training process of the preset liveness detection model, and at the time of the screen click operation. In this specification, the screen click object includes the training object. Specifically, when the sampled object at the time of the screen click operation is the same as the sampled object before the screen click operation, the screen click object is the training object. When the sampled object at the time of the screen click operation is different from the sampled object before the screen click operation, the screen click object and the training object are different people.

[0109] In this specification, server 130 can acquire a first image sample, screen click samples, and a second image sample through client 110. The first image sample includes facial image samples of the training object before performing a screen click operation. The screen click samples include screen click data generated during the screen click operation. The screen click samples may include at least one of the following: click location information, click time interval, click duration, and click force. The second image sample includes facial image samples of the screen click object during the screen click operation.

[0110] S520, the first image sample, the screen click sample and the second image sample are input into a preset liveness detection model to obtain the click point distribution prediction result, the click prediction result and the liveness prediction result.

[0111] Specifically, the preset liveness detection model may include a preset click point distribution generation network, a preset click pattern feature map extraction network, and a preset multimodal liveness detection network. In this specification, the preset click point distribution generation network, the preset click pattern feature map extraction network, and the preset multimodal liveness detection network are three neural networks with different functions in the preset liveness detection model. These networks can operate independently or collaboratively. The server 130 can input different data based on the functional characteristics of each of the three network structures to obtain corresponding output results.

[0112] Figure 11 A flowchart 520 illustrates a method for training a liveness detection model according to some embodiments of this specification.

[0113] In some embodiments, such as Figure 11 As shown, S520 may include:

[0114] S522, input the first image sample into the preset click point distribution generation network to obtain the click point distribution prediction result.

[0115] In this specification, server 130 can use a preset click point distribution generation network to process the input first image sample and predict the click point distribution prediction result. Specifically, the preset click point distribution generation network can include a preset facial feature extraction subnetwork and a preset distribution map prediction subnetwork. Server 130 can use the preset facial feature extraction subnetwork to extract features from the first image sample to obtain a facial feature map of the first image sample; and use the preset distribution map prediction subnetwork to classify the facial feature map to obtain the click point distribution prediction result.

[0116] Furthermore, the preset click point distribution generation network may also include a preset gaze pose prediction subnetwork. The server 130 can use the preset facial feature extraction subnetwork to extract features from the first image sample to obtain the facial feature map of the first image sample; and use the preset gaze pose prediction subnetwork to extract features from the facial feature map to obtain the gaze pose prediction result of the first image sample; then use the preset distribution map prediction subnetwork to classify and predict the facial feature map and the gaze pose prediction result to obtain the click point distribution prediction result.

[0117] It is important to note that server 130 uses a preset liveness detection model to process the first image sample, and the predicted click point distribution will serve as the basis for server 130 to randomly generate click point indication information. Since the screen click object subsequently needs to perform screen click operations under the guidance of the click point indication information, the screen click sample is the screen click data generated when the screen click object performs screen click operations under the guidance of the click point indication information.

[0118] S524, input the screen click sample into the preset click pattern feature map extraction network to obtain the click pattern feature map and the click prediction result.

[0119] Figure 12 A schematic diagram of the data flow of a preset click pattern feature map extraction network provided according to some embodiments of this specification is shown.

[0120] like Figure 12 As shown in this specification, the preset click pattern feature map extraction network includes a preset click feature extraction sub-network, a preset identity classification sub-network, and a preset click point distribution regression sub-network. The preset click feature extraction sub-network can be understood as a feature map encoder, while the preset identity classification sub-network and the preset click point distribution regression sub-network can both be understood as classifiers.

[0121] S524 may include: using the preset click feature extraction subnetwork to extract features from the screen click samples to obtain the click pattern feature map; using the preset identity classification subnetwork to predict the click pattern feature map to obtain an identity prediction result; and using the preset click point distribution regression subnetwork to predict the click pattern feature map to obtain a click point distribution regression prediction result, wherein the click prediction result includes the identity prediction result and the click point distribution regression prediction result.

[0122] Server 130 can use a preset click pattern feature map extraction network to extract features from the input screen click samples, obtaining a click pattern feature map, and then perform classification prediction based on the click pattern feature map to obtain the click prediction result. Specifically, the preset click pattern feature map extraction network can include a preset click feature extraction sub-network. Server 130 can use the preset click feature extraction sub-network to extract features from the screen click samples, obtaining a click pattern feature map. The click pattern feature map includes feature information corresponding to the click position, click time interval, click duration, and click intensity when the screen click object performs a screen click operation.

[0123] In this specification, the identity prediction result can be understood as the predicted personality type of the screen-clicking object. As mentioned earlier, different screen-clicking objects have different habits in performing screen-clicking operations, and the habitual characteristics of screen-clicking objects can reflect their personality traits to a certain extent. Furthermore, as mentioned earlier, the click pattern feature map includes the click location information, click time interval, click duration, and click intensity corresponding to the screen-clicking object's click operation. Therefore, the server 130 can use a preset identity classification sub-network to classify and predict the click pattern feature map to obtain the identity prediction result of the screen-clicking object. In this specification, the server 130 can classify the personality of the screen-clicking object into various types; for example, personality types can include kind, irritable, introverted, lively, gentle, etc.

[0124] The click distribution regression prediction result can be understood as the click probability of a screen click object performing a screen click operation when the screen click object is determined by the preset click pattern feature map extraction network based on the click pattern feature map. When the screen click object and the training object are the same person, and the screen click object performs screen click operations according to its inherent habits, the click distribution regression prediction result obtained by server 130 based on the screen click data should be consistent with the click distribution annotation result.

[0125] S526, input the click pattern feature map and the second image sample into the preset multimodal liveness detection network to obtain the liveness prediction result.

[0126] Figure 13 A schematic diagram of the data flow of a preset multimodal liveness detection network provided according to some embodiments of this specification is shown.

[0127] like Figure 13 As shown in this specification, the preset multimodal liveness detection network includes a preset facial feature encoding subnetwork, a preset click feature encoding subnetwork, and a preset liveness decision subnetwork.

[0128] S526 may include: using the preset facial feature encoding subnetwork to extract features from the second image sample to obtain a facial decision feature map; using the preset click feature encoding subnetwork to extract features from the click pattern feature map to obtain a click liveness feature map; and using the preset liveness decision subnetwork to predict the facial decision feature map and the click liveness feature map to obtain the liveness prediction result.

[0129] In this specification, server 130 can combine the feature information contained in both the facial decision feature map and the click liveness feature map to make a prediction and obtain a liveness prediction result. The liveness prediction result can be understood as the probability that the training object passes the current liveness detection.

[0130] In this specification, the preset liveness decision subnetwork includes a preset facial feature decision subnetwork, a preset click liveness feature decision subnetwork, and a preset fusion decision subnetwork. Furthermore, the step of predicting the facial decision feature map and the click liveness feature map based on the preset liveness decision subnetwork to obtain the liveness prediction result may include: predicting the facial decision feature map using the preset facial feature decision subnetwork to obtain a first liveness prediction result; predicting the click liveness feature map using the preset click liveness feature decision subnetwork to obtain a second liveness prediction result; and predicting the facial decision feature map and the click liveness feature map using the preset fusion decision subnetwork to obtain a third liveness prediction result, wherein the liveness prediction result includes the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result.

[0131] In this specification, the first liveness prediction result can be understood as the liveness category predicted by server 130 based on the facial features of the screen-clicked object; different screen-clicked objects correspond to different liveness categories. The second liveness prediction result can be understood as the liveness prediction result predicted by server 130 based on the habitual features reflected in the screen-clicking operation of the screen-clicked object. The third liveness prediction result can be understood as the liveness prediction result predicted by server 130 after fusing the facial decision feature map and the click liveness feature map, based on the habitual features reflected in the screen-clicking operation of the screen-clicked object and the facial features of the screen-clicked object. During the fusion of the facial decision feature map and the click liveness feature map, server 130 can prioritize the facial decision feature map and use the click liveness feature map as a supplement. Specifically, server 130 can adjust the weights of the facial decision feature map and the click liveness feature map, making the weight of the facial decision feature map greater than the weight of the click liveness feature map, so that the facial decision feature map has a greater influence on the third liveness prediction result.

[0132] S530, the preset liveness detection model is trained based on the click point distribution prediction result, the click prediction result, and the liveness prediction result to obtain the trained liveness detection model.

[0133] In this specification, the click point distribution prediction result, click prediction result, and liveness prediction result are the outputs of different sub-networks in the preset liveness detection model. Therefore, the training of the preset liveness detection model by the server 130 can be understood as follows: the server 130 trains the preset click point distribution generation network, the preset click pattern feature map extraction network, and the preset multimodal liveness detection network contained in the preset liveness detection model, respectively. The trained click point distribution generation network, the trained click pattern feature map extraction network, and the trained multimodal liveness detection network together constitute the trained liveness detection model.

[0134] Figure 14 A flowchart 530 is shown of a method for training a liveness detection model according to some embodiments of this specification.

[0135] In some embodiments, such as Figure 14 As shown, S530 may include:

[0136] S532, the preset click point distribution generation network is trained based on the click point distribution prediction results to obtain the trained click point distribution generation network. In this specification, the click point distribution prediction results are used as the output results of the preset click point distribution generation network. The server 130 can use the click point distribution prediction results as the training basis for the preset click point distribution generation network to update the network parameters of the preset click point distribution generation network, thereby obtaining the trained click point distribution generation network.

[0137] Specifically, S532 may include: obtaining the click point distribution annotation result of the training object; comparing the click point distribution prediction result with the click point distribution annotation result to obtain first click point distribution loss information; and converging the preset click point distribution generation network based on the first click point distribution loss information to obtain the trained click point distribution generation network.

[0138] In this specification, the click point distribution annotation result can be understood as the click probability of a training subject clicking on various click areas on the screen when performing screen click operations without external guidance. There are multiple ways to obtain the click point distribution annotation result. For example, server 130 can display clickable points available to the training subject on the client 110 screen. The client 110 screen can display a certain number of independent areas (each independent area corresponds to one click point), and the training subject can click according to requirements (e.g., a specified number of clicks). Taking the client 110 screen displaying 9 click points and requiring the training subject to input 6 click points (the 6 click points can be the same or different) as an example: The client 110 screen displays all 9 click points. During the first input of click points, the training subject can randomly click 6 of them. To improve the stability of the click point distribution annotation result input by the training subject, the training subject can input 6 click points a second time after the first input. The requirements for the second input click points are the same as those for the first input click points. The distribution of the 6 click points input by the training subject in the second input may be different from the distribution of the 6 click points input in the first input. Afterward, the training subject can perform a third, fourth, or more click point input operations. The server 130 can pre-set the number of click points that the training subject can input.

[0139] After the training subject completes the prescribed number of inputs, the server 130 can determine the click point distribution annotation result based on the click point input. For example, the server 130 can select the most representative result as the click point distribution annotation result. Specifically, the server 130 can select the click point distribution corresponding to the shortest time taken for the training subject to input 6 click points as the click point distribution annotation result.

[0140] It should be noted that during the training phase of the liveness detection model, the number and arrangement of click points that can be displayed on the client 110 screen should be consistent with those during the application phase of the liveness detection model.

[0141] In this specification, the first click point distribution loss information can be understood as the loss information formed by the difference between the click point distribution annotation result and the click point distribution prediction result. The first click point classification loss information is used to ensure that the click point distribution prediction result predicted by the preset click point distribution generation network is consistent with the click point distribution annotation result.

[0142] In this specification, after determining the first click point distribution loss information, server 130 can converge the preset click point distribution generation network based on the first click point distribution loss information, thereby obtaining the trained click point distribution generation network. Server 130 can use various methods to converge the preset click point distribution generation network based on the first click point distribution loss information. For example, server 130 can use gradient descent to update the network parameters of the preset click point distribution generation network based on the first click point distribution loss information; then, server 130 can return to the step of inputting the first image sample into the preset click point distribution generation network until the preset click point distribution generation network converges, thereby obtaining the trained click point distribution generation network. Server 130 can also use other parameter update algorithms to update the network parameters of the preset click point distribution generation network based on the first click point distribution loss information; then, server 130 can return to the step of inputting the first image sample into the preset click point distribution generation network until the preset click point distribution generation network converges, thereby obtaining the trained click point distribution generation network.

[0143] S534, the preset click pattern feature map extraction network is trained based on the click prediction results to obtain the trained click pattern feature map extraction network. In this specification, both the identity prediction results and the click point distribution regression prediction results are used as the output results of the preset click pattern feature map extraction network. The server 130 can use both the identity prediction results and the click point distribution regression prediction results as the training basis for the preset click pattern feature map extraction network to update the network parameters of the preset click pattern feature map extraction network, thereby obtaining the trained click pattern feature map extraction network.

[0144] Specifically, training the preset click pattern feature map extraction network based on the click prediction results to obtain the trained click pattern feature map extraction network may include: obtaining the identity labeling results of the training object and comparing the identity prediction results with the identity labeling results to obtain identity classification loss information; obtaining the click point distribution labeling results of the training object and comparing the click point distribution regression prediction results with the click point distribution labeling results to obtain second click point distribution loss information; and converging the preset click pattern feature map extraction network based on the identity classification loss information and the second click point distribution loss information to obtain the trained click pattern feature map extraction network.

[0145] In this specification, the identity labeling result can be understood as the personality type labeled by server 130 for screen click objects. Identity classification loss information can be understood as the loss information formed by the difference between the identity labeling result and the identity prediction result. The identity classification loss information is used to ensure that the identity prediction result predicted by the preset click pattern feature map extraction network remains consistent with the actual identity labeling result.

[0146] In this specification, the click point distribution annotation result can be understood as the click probability of a screen click object performing a screen click operation on various click areas on the screen without external guidance. The second click point distribution loss information can be understood as the loss information formed by the difference between the click point distribution annotation result and the click point distribution regression prediction result. This second click point distribution loss information is used to ensure that the click point distribution regression prediction result predicted by the preset click pattern feature map extraction network remains consistent with the actual click point distribution annotation result.

[0147] In this specification, after determining the identity classification loss information and the second click point distribution loss information, server 130 can converge the preset click pattern feature map extraction network based on the identity classification loss information and the second click point distribution loss information, thereby obtaining the trained click pattern feature map extraction network. Server 130 can converge the preset click pattern feature map extraction network based on the identity classification loss information and the second click point distribution loss information in various ways. For example, server 130 can fuse the identity classification loss information and the second click point distribution loss information to obtain comprehensive click loss information, and then converge the preset click pattern feature map extraction network based on comprehensive click loss information. Server 130 can fuse the identity classification loss information and the second click point distribution loss information in various ways. For example, server 130 can directly add the identity classification loss information and the second click point distribution loss information to obtain comprehensive click loss information. Server 130 can also obtain the loss weights of identity classification loss information and second click point distribution loss information respectively, and weight the identity classification loss information and second click point distribution loss information based on their respective loss weights. Then, the weighted identity classification loss information and the weighted second click point distribution loss information are added together to obtain the comprehensive click loss information.

[0148] After obtaining the comprehensive click loss information, server 130 can use gradient descent to update the network parameters of the preset click pattern feature map extraction network based on the comprehensive click loss information. Then, server 130 can return to the step of inputting screen click samples into the preset click pattern feature map extraction network until the preset click pattern feature map extraction network converges, thus obtaining the trained click pattern feature map extraction network. Server 130 can also use other parameter update algorithms to update the network parameters of the preset click pattern feature map extraction network based on the comprehensive click loss information; then, server 130 can return to the step of inputting screen click samples into the preset click pattern feature map extraction network until the preset click pattern feature map extraction network converges, thus obtaining the trained click pattern feature map extraction network.

[0149] S536, the preset multimodal liveness detection network is trained based on the liveness prediction results to obtain the trained multimodal liveness detection network. In this specification, the liveness prediction results are used as the output of the preset multimodal liveness detection network. The server 130 can use the liveness prediction results as the training basis for the preset multimodal liveness detection network to update the network parameters of the preset multimodal liveness detection network, thereby obtaining the trained multimodal liveness detection network.

[0150] Specifically, S536 may include: determining the liveness classification loss information of the multimodal liveness detection network based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result; comparing the first liveness prediction result and the third liveness prediction result to obtain first consistency prediction loss information; comparing the second liveness prediction result and the third liveness prediction result to obtain second consistency prediction loss information; and converging the preset multimodal liveness detection network based on the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information to obtain the trained multimodal liveness detection network.

[0151] In this specification, the liveness classification loss information can be understood as the loss information formed by the difference between the predicted liveness classification result and the actual liveness annotation result. The liveness classification loss information may include first liveness classification loss information, second liveness classification loss information, and third liveness classification loss information. Determining the liveness classification loss information of the multimodal liveness detection network based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result includes: obtaining a first liveness annotation result of the training object based on facial features, and comparing the first liveness prediction result with the first liveness annotation result to obtain the first liveness classification loss information; obtaining a second liveness annotation result of the training object based on click point operations, and comparing the second liveness prediction result with the second liveness annotation result to obtain the second liveness classification loss information; and obtaining a third liveness annotation result of the training object based on the facial features and the click point operations, and comparing the third liveness prediction result with the third liveness annotation result to obtain the third liveness classification loss information.

[0152] In this specification, the first liveness labeling result can be understood as the liveness category pre-labeled by server 130 based on the facial features of the training object; different training objects correspond to different liveness categories. The first liveness classification loss information can be understood as the loss information formed by the difference between the predicted first liveness prediction result and the actual first liveness labeling result. The second labeling result can be understood as the habit category pre-labeled by server 130 based on the habitual features reflected when the training object performs click operations; different training objects may correspond to different habitual categories. The second liveness classification loss information can be understood as the loss information formed by the difference between the predicted second liveness prediction result and the actual second liveness labeling result. The first liveness labeling result can be understood as the liveness category pre-labeled by server 130 after integrating the facial features of the training object and the habitual features reflected when the training object performs click operations; different training objects correspond to different liveness categories. The third liveness classification loss information can be understood as the loss information formed by the difference between the predicted third liveness prediction result and the actual third liveness labeling result.

[0153] In this specification, server 130 can also compare the first liveness prediction result and the third liveness prediction result to obtain first consistency prediction loss information. The first consistency prediction loss information can be understood as the loss information formed by the difference between the first liveness prediction result and the third liveness prediction result. Furthermore, server 130 can also compare the second liveness prediction result and the third liveness prediction result to obtain second consistency prediction loss information. The second consistency prediction loss information can be understood as the loss information formed by the difference between the second liveness prediction result and the third liveness prediction result.

[0154] In this specification, after determining the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information, the server 130 can converge a preset multimodal liveness detection network based on these loss information to obtain a trained modal liveness detection network. The server 130 can converge the preset multimodal liveness detection network using various methods. For example, the server 130 can fuse the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information to obtain multimodal comprehensive loss information, and then converge the preset multimodal liveness detection network based on this comprehensive loss information.

[0155] Server 130 can fuse the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information in several ways. For example, server 130 can directly add the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information to obtain multimodal comprehensive loss information. Server 130 can also obtain the loss weights of the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information respectively, and weight the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information based on their respective loss weights, and then add the weighted liveness classification loss information, the weighted first consistency prediction loss information, and the weighted second consistency prediction loss information to obtain multimodal comprehensive loss information.

[0156] After obtaining the multimodal comprehensive loss information, server 130 can use gradient descent to update the network parameters of the preset multimodal liveness detection network based on the multimodal comprehensive loss information. Then, server 130 can return to the step of inputting the click pattern feature map and the second image sample into the preset multimodal liveness detection network until the preset multimodal liveness detection network converges, thus obtaining the trained multimodal liveness detection network. Server 130 can also use other parameter update algorithms to update the network parameters of the preset multimodal liveness detection network based on the multimodal comprehensive loss information; then, server 130 can return to the step of inputting the click pattern feature map and the second image sample into the preset multimodal liveness detection network until the preset multimodal liveness detection network converges, thus obtaining the trained multimodal liveness detection network.

[0157] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification is intended to encompass various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0158] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0159] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0160] It should be understood that in the foregoing description of the embodiments of this specification, for the purpose of simplifying the description and to aid in understanding a feature, various features are sometimes combined in a single embodiment, drawing, or description thereof. Alternatively, various features may be distributed across multiple embodiments of this specification. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may extract some features as individual embodiments for understanding. That is, the embodiments in this specification can also be understood as an integration of multiple sub-embodiments. It is also valid when each sub-embodiment contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A method for detecting liveness, comprising: A first facial image of the target object is obtained, and the first facial image is input into a liveness detection model to obtain the predicted click point distribution of the target object. Based on the click point distribution prediction results, click point indication information of the target object is displayed, wherein the click point distribution prediction results reflect the click probability of the target object on each click area on the screen, and the click point indication information is used to guide the target object to perform screen click operations; Obtain screen click data of the screen click object and a second facial image of the screen click object. The screen click data reflects the habitual characteristics of the screen click object during the screen click operation according to the click point indication information, and the second facial image reflects the identity characteristics of the screen click object during the screen click operation. Based on the second facial image and the screen click data, a liveness detection result is obtained to determine whether the screen click object is the target object; and Output the liveness detection results.

2. The liveness detection method as described in claim 1, wherein, The liveness detection model includes a click point distribution generation network; as well as The step of inputting the first facial image into the liveness detection model to obtain the predicted click point distribution result of the target object includes: The click point distribution generation network is used to predict the first facial image to obtain the click point distribution prediction result of the target object.

3. The liveness detection method as described in claim 2, wherein, The click point distribution generation network includes a facial feature extraction subnetwork and a distribution map prediction subnetwork; as well as The step of using the click point distribution generation network to predict the first facial image and obtaining the click point distribution prediction result of the target object includes: The facial feature extraction sub-network is used to extract features from the first facial image to obtain a facial feature map, and The distribution map prediction subnetwork is used to predict the facial feature map to obtain the click point distribution prediction result.

4. The live detection method as described in claim 3, wherein, The click point distribution generation network also includes a gaze pose prediction subnetwork; as well as The step of using the distribution map prediction subnetwork to predict the facial feature map to obtain the click point distribution prediction result includes: The gaze pose prediction sub-network is used to extract features from the facial feature map to obtain the gaze pose prediction result. The distribution map prediction subnetwork is used to predict the facial feature map and the gaze pose prediction results to obtain the click point distribution prediction results.

5. The liveness detection method as described in claim 1, wherein, The screen click data includes at least one of the following: click location information, click time interval, click duration, and click force when the screen click object performs the screen click operation.

6. The liveness detection method as described in claim 1, wherein, The liveness detection model also includes a click pattern feature map extraction network and a multimodal liveness detection network; as well as The step of determining whether the screen click object is the target object based on the second facial image and the screen click data is a liveness detection result, including: The click pattern feature map extraction network is used to extract features from the screen click data to obtain the click pattern feature map of the screen click object. The multimodal liveness detection network is used to predict the second facial image and the click pattern feature map to obtain the liveness detection result of the target object.

7. The liveness detection method as described in claim 6, wherein, The multimodal liveness detection network includes a facial feature encoding subnetwork, a click feature encoding subnetwork, and a fusion decision subnetwork; and The step of using the multimodal liveness detection network to predict the second facial image and the click pattern feature map to obtain the liveness detection result of the target object includes: The facial feature encoding subnetwork is used to extract features from the second facial image to obtain a facial decision feature map. The click feature encoding subnetwork is used to extract features from the click pattern feature map to obtain a click liveness feature map. The fusion decision sub-network is used to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object.

8. The liveness detection method as described in claim 7, wherein, The step of using the fusion decision sub-network to predict the facial decision feature map and the click liveness feature map to obtain the liveness detection result of the target object includes: The facial decision feature map and the click liveness feature map are fused to obtain a fused feature map; and The fusion decision sub-network is used to predict the fusion feature map to obtain the liveness detection result of the target object.

9. A liveness detection system, comprising: At least one storage medium, including at least one instruction set, for implementation analysis of the liveness detection method; as well as At least one processor is communicatively connected to the at least one storage medium. When the system is running, the at least one processor reads the at least one instruction set and executes the method of any one of claims 1-8 according to the instructions of the at least one instruction set.

10. A training method for a liveness detection model, comprising: Obtain a first image sample of the training object, a screen click sample of the screen click object, and a second image sample of the screen click object, wherein the screen click object is the sampling object when performing a screen click operation; The first image sample is input into a preset liveness detection model to obtain a click point distribution prediction result; the screen click sample is input into the preset liveness detection model to obtain a click prediction result and a click pattern feature map; the second image sample is input into the preset liveness detection model, and based on the click pattern feature map, a liveness prediction result is obtained; wherein, the click prediction result includes an identity prediction result and a click point distribution regression prediction result; and The preset liveness detection model is trained based on the click point distribution prediction results, the click prediction results, and the liveness prediction results to obtain the trained liveness detection model.

11. The training method as described in claim 10, wherein, The preset liveness detection model includes a preset click point distribution generation network, a preset click pattern feature map extraction network, and a preset multimodal liveness detection network; as well as The first image sample is input into a preset liveness detection model to obtain the click point distribution prediction result; The screen click samples are input into a preset liveness detection model to obtain click prediction results and click pattern feature maps; The second image sample is input into a preset liveness detection model, and a liveness prediction result is obtained based on the click pattern feature map; wherein, the click prediction result includes an identity prediction result and a click point distribution regression prediction result, including: The first image sample is input into the preset click point distribution generation network to obtain the click point distribution prediction result. The screen click samples are input into the preset click pattern feature map extraction network to obtain the click pattern feature map and the click prediction result. The click pattern feature map and the second image sample are input into the preset multimodal liveness detection network to obtain the liveness prediction result.

12. The training method as described in claim 11, wherein, The preset click pattern feature map extraction network includes a preset click feature extraction sub-network, a preset identity classification sub-network, and a preset click point distribution regression sub-network; as well as The step of inputting the screen click samples into the preset click pattern feature map extraction network to obtain the click pattern feature map and the click prediction result includes: The preset click feature extraction subnetwork is used to extract features from the screen click samples to obtain the click pattern feature map. The preset identity classification subnetwork is used to predict the click pattern feature map to obtain the identity prediction result. The click pattern feature map is predicted using the preset click point distribution regression sub-network to obtain the click point distribution regression prediction result.

13. The training method as described in claim 11, wherein, The preset multimodal liveness detection network includes a preset facial feature encoding subnetwork, a preset click feature encoding subnetwork, and a preset liveness decision subnetwork; as well as The step of inputting the click pattern feature map and the second image sample into the preset multimodal liveness detection network to obtain the liveness prediction result includes: The preset facial feature encoding subnetwork is used to extract features from the second image sample to obtain a facial decision feature map. The preset click feature encoding subnetwork is used to extract features from the click pattern feature map to obtain a click liveness feature map. Based on the preset liveness decision sub-network, the facial decision feature map and the click liveness feature map are predicted to obtain the liveness prediction result.

14. The training method as described in claim 13, wherein, The preset liveness decision subnetwork includes a preset facial feature decision subnetwork, a preset click liveness feature decision subnetwork, and a preset fusion decision subnetwork; as well as The prediction of the facial decision feature map and the click liveness feature map based on the preset liveness decision sub-network to obtain the liveness prediction result includes: The facial decision feature map is predicted using the preset facial feature decision sub-network to obtain the first liveness prediction result. The preset click-to-liveness feature decision sub-network is used to predict the click-to-liveness feature map to obtain a second liveness prediction result. The preset fusion decision sub-network is used to predict the facial decision feature map and the click liveness feature map to obtain a third liveness prediction result. The liveness prediction results include the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result.

15. The training method as described in claim 14, wherein, The step of training the preset liveness detection model based on the click point distribution prediction result, the click prediction result, and the liveness prediction result includes: The preset click point distribution generation network is trained based on the click point distribution prediction results to obtain the trained click point distribution generation network. The preset click pattern feature map extraction network is trained based on the click prediction results to obtain the trained click pattern feature map extraction network; and The preset multimodal liveness detection network is trained based on the liveness prediction results to obtain the trained multimodal liveness detection network.

16. The training method as described in claim 15, wherein, The step of training the preset click point distribution generation network based on the click point distribution prediction results to obtain the trained click point distribution generation network includes: Obtain the click point distribution annotation results of the training object; The click point distribution prediction result is compared with the click point distribution annotation result to obtain the first click point distribution loss information; and The preset click point distribution generation network is converged based on the first click point distribution loss information to obtain the trained click point distribution generation network.

17. The training method as described in claim 15, wherein training the preset click pattern feature map extraction network based on the click prediction result to obtain the trained click pattern feature map extraction network includes: Obtain the identity labeling results of the training objects, and compare the identity prediction results with the identity labeling results to obtain identity classification loss information; Obtain the click point distribution annotation results of the training object, and compare the click point distribution regression prediction results with the click point distribution annotation results to obtain the second click point distribution loss information; and The preset click pattern feature map extraction network is converged based on the identity classification loss information and the second click point distribution loss information to obtain the trained click pattern feature map extraction network.

18. The training method as described in claim 15, wherein, The step of training the preset multimodal liveness detection network based on the liveness prediction results to obtain the trained multimodal liveness detection network includes: The liveness classification loss information of the multimodal liveness detection network is determined based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction result. The first liveness prediction result and the third liveness prediction result are compared to obtain the first consistency prediction loss information; The second liveness prediction result and the third liveness prediction result are compared to obtain the second consistency prediction loss information; and The preset multimodal liveness detection network is converged based on the liveness classification loss information, the first consistency prediction loss information, and the second consistency prediction loss information to obtain the trained multimodal liveness detection network.

19. The training method as described in claim 18, wherein, The liveness classification loss information includes first liveness classification loss information, second liveness classification loss information, and third liveness classification loss information; and The step of determining the liveness classification loss information of the multimodal liveness detection network based on the first liveness prediction result, the second liveness prediction result, and the third liveness prediction includes: The first liveness labeling result based on facial features of the training object is obtained, and the first liveness prediction result is compared with the first liveness labeling result to obtain the first liveness classification loss information. Obtain the second liveness labeling result of the training object based on the click point operation, and compare the second liveness prediction result with the second liveness labeling result to obtain the second liveness classification loss information, and Obtain the third liveness labeling result of the training object based on the facial features and the click point operation, and compare the third liveness prediction result with the third liveness labeling result to obtain the third liveness classification loss information.

20. A training system for a liveness detection model, comprising: At least one storage medium, including at least one instruction set, for implementation analysis of training methods for a liveness detection model; as well as At least one processor is communicatively connected to the at least one storage medium. When the system is running, the at least one processor reads the at least one instruction set and executes the method of any one of claims 10-19 according to the instructions of the at least one instruction set.

Images