WO2021051510A1

WO2021051510A1 - Method and apparatus for generating face image, computer device, and storage medium

Info

Publication number: WO2021051510A1
Application number: PCT/CN2019/116707
Authority: WO
Inventors: 徐玲玲
Original assignee: 平安科技（深圳）有限公司
Priority date: 2019-09-19
Filing date: 2019-11-08
Publication date: 2021-03-25
Also published as: CN110781738A

Abstract

Disclosed are a method and apparatus for generating a face image, a computer device, and a storage medium. The method for generating a face image comprises: simultaneously inputting an original face image and an angle vector of a target face image into a pre-trained face image generation network; converting, by the pre-trained face image generation network, the original face image into a first target face image corresponding to a first angle vector according to a preset conversion mode, the first angle vector being comprised in angle vectors corresponding to all target face images; the first target face image being comprised in all target face images; evaluating, by a preset determination network, whether the first angle vector matches the first target face image; and if the first angle vector matches the first target face image, outputting the first target face image. The synthesis of the target face image corresponding to an angle is achieved by simultaneously inputting the original face image and the angle vector into the face image generation network and acting the angle vector on the original face image.

Description

Method, device, computer equipment and storage medium for generating face image

This application claims the priority of the Chinese patent application filed with the Chinese Patent Office on September 19, 2019, the application number is 2019108866185, and the invention title is "Methods, Apparatus, Computer Equipment and Storage Media for Generating Face Images", and its entire contents Incorporated in this application by reference.

Technical field

This application relates to the field of computers, in particular to methods, devices, computer equipment and storage media for generating facial images.

Background technique

Face recognition has been widely used in robotics and other disciplines. It is of great significance for automatic identification and automatic discrimination of humans. However, when passing face recognition verification, each user usually only collects a front view image. The face recognition neural network can only generate one type of picture, and cannot generate pictures of multiple angle types at the same time. This makes the application of the real face recognition system have recognition defects, which leads to a sharp drop in the recognition rate and affects normal use.

technical problem

The main purpose of this application is to provide a method for generating a face image, which aims to solve the technical problem that the existing face recognition neural network cannot generate pictures of multiple angle types at the same time.

Technical solutions

This application proposes a method for generating a face image, including:

Input the angle vectors of the original face image and the target face image into the pre-trained face image generation network at the same time, where the angle vector of the target face image includes at least one;

Through the pre-trained face image generation network, the original face image is converted into the first target face image corresponding to the first angle vector according to the preset conversion method, where the first angle vector is included in all target face images corresponding to In the angle vector, the first target face image is included in all target face images;

Evaluate whether the first angle vector matches the first target face image through a preset discrimination network;

If it matches, output the first target face image.

This application also provides a device for generating a face image, including:

The first input module is used to input the original face image and the angle vector of the target face image into the pre-trained face image generation network at the same time, wherein the angle vector of the target face image includes at least one;

The conversion module is used to convert the original face image into the first target face image corresponding to the first angle vector through the pre-trained face image generation network according to the preset conversion method, where the first angle vector is included in all targets In the angle vector corresponding to the face image, the first target face image is included in all target face images;

The evaluation module is used to evaluate whether the first angle vector matches the first target face image through a preset discrimination network;

The output module is used to output the first target face image if it matches.

The present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when the computer program is executed.

The present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned method are realized.

Beneficial effect

In this application, the original face image and the angle vector are input into the face image generation network together, and the angle vector is applied to the original face image to realize the synthesis of the target face image corresponding to the angle. The face image generation network of the present application is formed by multi-user and multi-angle face image training, and carries multiple angle images corresponding to the generation template, which can support an original face image and input multiple angle vectors to the face. In the image generation network, target face images of multiple angles corresponding to the original face image are generated at the same time. By connecting the generation network and the discrimination network, the reliability of the generation network is further verified by angle classification of the face images output by the generation network. If the face image corresponding to the angle is classified into the correct angle domain by the discrimination network, it indicates The reliability of the generation network is high, and the reliability of the discrimination network is also high. If the classification is incorrect, but the generation network is reliable after the first correction loss function analysis, it indicates that the discrimination network needs to be corrected, and the second correction loss function Ld is used for correction.

Description of the drawings

Fig. 1 is a schematic flowchart of a method for generating a face image according to an embodiment of the present application;

Fig. 2 is a schematic structural diagram of an apparatus for generating a face image according to an embodiment of the present application;

Fig. 3 is a schematic diagram of the internal structure of a computer device according to an embodiment of the present application.

The best mode of the present invention

1, a method for generating a face image according to an embodiment of the present application includes:

S1: The angle vectors of the original face image and the target face image are simultaneously input into the pre-trained face image generation network, where the angle vector of the target face image includes at least one.

S2: Through the pre-trained face image generation network, the original face image is converted into the first target face image corresponding to the first angle vector according to the preset conversion method, where the first angle vector is included in all target face images In the corresponding angle vector, the first target face image is included in all target face images.

S3: Evaluate whether the first angle vector matches the first target face image through a preset discrimination network.

S4: If it matches, output the first target face image.

In this embodiment, the original face image and the angle vector of the target face image are input into the face image generation network together, and the angle vector is applied to the original face image to realize the target face image corresponding to the angle vector Synthesis. The face image generation network of this embodiment is formed by multi-user and multi-angle face image training, and carries multiple angle images corresponding to the generation template, which can support an original face image and input multiple angle vectors together. In the face image generation network, the target face images corresponding to the original face images after the conversion of each angle vector are generated and output at the same time.

Further, the step S2 of converting the original face image into the first target face image corresponding to the first angle vector according to the preset conversion method through the pre-trained face image generation network includes:

S21: Cut out an image of the head region of the face from the original face image.

S22: Locate the feature points corresponding to the designated organs of the face according to the feature point positioning model, where the designated organs include at least one type.

S23: Obtain coordinate data of feature points corresponding to each designated organ.

S24: Map each coordinate data one-to-one to the feature point template corresponding to the first angle vector to obtain the face area corresponding to the first angle vector.

S25: Form a first target face image according to the face area corresponding to the first angle vector.

In this embodiment, in the process of synthesizing the target face image corresponding to the angle, in order to improve the accurate matching between the generated target face image and the original face image, the head image of the face is enlarged and displayed to form a face. The head region image is recognized and located according to the positioning model, and the coordinate data of the feature point is mapped to the feature point template corresponding to the angle vector according to the angle vector to obtain the target face image. The above-mentioned feature points of the human face include at least five feature points, the feature points corresponding to the two eyes, the feature points corresponding to the corners of the two mouths, and the feature point corresponding to the nose. In this embodiment, after obtaining the coordinate data of the five feature points in the original face image, each coordinate is represented by (x, y), and then according to the opencv affine transformation, the coordinates of the five feature points obtained are The data is mapped to the feature point template corresponding to the pre-stored angle vector, and the face area corresponding to the angle vector is obtained to form the target face image. According to the face detection model (such as mtcnn), five feature points of the face can be obtained, and the coordinate data corresponding to the five feature points can be determined.

Further, the step S3 of evaluating whether the first angle vector matches the first target face image through a preset discrimination network includes:

S31: Input the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network.

S31: Generate a second target face image according to the inverse angle of the first angle vector and the first target face image.

S33: Determine whether the second target face image is the same as the original face image.

S34: If they are the same, determine that the first angle vector matches the first target face image.

This embodiment verifies the reliability of the pre-trained face image generation network by using the inverse angle of the corresponding angle of the target face image and the synthesized target face image as the input of the face image generation network, and whether to output the original face image If the inverse angle of the corresponding angle between the target face image and the synthesized target face image is used as the input of the face image generation network, and the original face image can be output, it indicates that the face image generation network is reliable. The above-mentioned synthetic target face image corresponding angle is c, and the inverse angle of the synthetic target face image corresponding angle is -c.

Further, after step S33 of determining whether the second target face image is the same as the original face image, the method includes:

S35: If the second target face image is different from the original face image, determine the difference information between the second target face image and the original face image.

S36: Generate a first correction loss function according to the difference information, where the first correction loss function is Lg=Ladv+Lfcls+Lrec, Ladv=logD(x)+log(1-D(G(x,c)), Lfcls =-logD(G(x,c)), Lrec=||xG(G(x,c),c')||, G represents the face image generation network, G(x,c) represents the input image x, When generating the corresponding face image when the angle is c, D(x) represents the probability distribution of the domain label calculated by D on the real image, D represents the discriminant network function, and c'=-c.

S37: Correct the parameters in the pre-trained face image generation network according to the first correction loss function.

In this embodiment, if the second target face image is different from the original face image, or the similarity between the second target face image and the original face image is less than the preset threshold, it means that the reliability of the face image generation network is not To meet the usage requirements, the first correction loss function is Lg, and the face image generation network G is corrected by the first correction loss function. G(x,c) means to generate a face image with an angle of c, G(G(x,c),c') means to generate an image with an angle of c as input again to generate the person corresponding to its original angle c' Face image, c'=-c. Lfcls is used to optimize the generation network G, that is, the face image generated by the generation network G can be classified into the target angle domain c by the discrimination network D. Lrec is to ensure that the original face image x is transformed in the angle domain c, and then transformed back to the face image corresponding to the original angle domain c', which is consistent with the original face image.

Further, after step S4 of outputting the first target face image, the method includes:

S41: Determine whether the angle classification corresponding to each first target face image is correct.

S42: If not, use the second correction loss function to correct the discrimination network, where the second correction loss function Ld=-Ladv+Lfcls, Ladv=logD(x)+log(1-D(G(x,c)) , Lfcls=-logD(G(x,c)), G(x,c) represents the input image x, when the generated angle is c, the corresponding face image, D(x) represents the domain calculated by D on the real image Label probability distribution, D represents the discriminant network function.

In this embodiment, the generation network is connected with the discrimination network, and the face image output by the generation network is classified by angle to further verify the reliability of the generation network. If the face image corresponding to the angle is classified to the correct angle by the discrimination network Domain indicates that the reliability of the generated network is high, and the reliability of the discrimination network is also high. If the classification is incorrect, but the generation network is reliable after the first correction loss function analysis, it indicates that the discrimination network needs to be corrected, and the second correction loss function Ld is used for correction. In this embodiment, by comparing the magnitude of the angle vector of the first target face image relative to the original face, whether it is within the corresponding angle classification range, and if it is within the range, it indicates that the classification is correct.

Further, inputting the angle vectors of the original face image and the target face image into the pre-trained face image generation network at the same time before step S1 includes:

S11: Obtain face images of a specified number of users, where the face images of the same user include at least two angle images corresponding to the specified angles.

S12: The specified angle corresponding to the angle image is mapped to a vector value of the specified length.

S13: Combine the angle image and the specified length vector value corresponding to the specified angle to form a training sample.

S14: Input the training samples into the neural network for training, and obtain a classification model that can classify the image domain corresponding to the specified angle.

S15: Use the classification model as a pre-trained face image generation network.

In this embodiment, when training the face image generation network, the parameters of the angle vector are added at the same time to generate face image generation templates corresponding to different angle domains. According to different angles, the face images are classified, the same angle is put into the same classification, and each angle is represented by an angle vector with a unique length. The specified number is greater than or equal to 1000 users, and each user corresponds to face images of multiple angles. The specified angles include, for example, -90, -60, -45, -30, 0, 30, 45, 60, 90 and other angles.

Further, the step S13 of combining the angle image and the vector value of the specified length corresponding to the specified angle into a training sample includes:

S131: Obtain the first pixel value corresponding to the facial organ in the first angle image, where the first angle image is included in all angle images.

S132: Obtain a standard face image corresponding to the first specified angle.

S133: Correct the first pixel value according to the standard pixel value corresponding to each face organ in the standard face image to obtain the corrected first pixel value.

S134: Combine the second angle image carrying the corrected first pixel value and the first vector value corresponding to the first angle into a corrected training sample.

In this embodiment, the training samples are corrected by positioning to improve the accuracy of the training samples. In this embodiment, the face image is centered by cropping, and then the pixel values of the feature points corresponding to the facial organs are corrected. For example, if the standard face image is 256*256, the standard pixel values of the eyes are 58*58, and 158*58. By correcting the first pixel value corresponding to the facial organ in the current user’s first angle image, the correction is aligned to The standard pixel value of the standard face image corresponding to the first specified angle, and the feature points corresponding to other face organs are correspondingly corrected in the above process to reduce the degree of sample difference, the impact on the training accuracy of the face image generation network, and improve the generation network Synchronously output the accuracy of face images corresponding to multiple angles.

Referring to FIG. 2, the apparatus for generating a face image according to an embodiment of the present application includes:

The first input module 1 is configured to simultaneously input the angle vectors of the original face image and the target face image into a pre-trained face image generation network, where the angle vector of the target face image includes at least one.

The conversion module 2 is used to convert the original face image into a first target face image corresponding to a first angle vector through a pre-trained face image generation network according to a preset conversion method, where the first angle vector is included in all In the angle vector corresponding to the target face image, the first target face image is included in all target face images.

The evaluation module 3 is used to evaluate whether the first angle vector matches the first target face image through a preset discrimination network.

The output module 4 is configured to output the first target face image if it matches.

Further, the conversion module 2 includes:

The cropping unit is used to crop the image of the head region of the face from the original face image.

The positioning unit is used for locating the feature points corresponding to the designated organs of the face according to the feature point positioning model, wherein the designated organs include at least one type.

The first acquiring unit is used to acquire the coordinate data of the characteristic points corresponding to each designated organ.

The mapping unit is used to map each coordinate data one-to-one to the feature point template corresponding to the first angle vector to obtain the face area corresponding to the first angle vector.

The forming unit is used to form the first target face image according to the face area corresponding to the first angle vector.

Further, the evaluation module 3 includes:

The input unit is used to input the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network.

The first generating unit is configured to generate a second target face image according to the inverse angle of the first angle vector and the first target face image.

The judging unit is used to judge whether the second target face image is the same as the original face image.

The determining unit is configured to determine that the first angle vector matches the first target face image if they are the same.

This embodiment verifies the reliability of the pre-trained face image generation network by using the inverse angle of the corresponding angle of the target face image and the synthesized target face image as the input of the face image generation network, and whether to output the original face image If the inverse angle of the corresponding angle between the target face image and the synthesized target face image is used as the input of the face image generation network, and the original face image can be output, it indicates that the face image generation network is reliable. The above-mentioned synthetic target face image corresponding angle is c, then the inverse angle of the synthetic target face image corresponding angle is -c.

Further, the evaluation module 3 includes:

The determining unit is configured to determine the difference information between the second target face image and the original face image if the second target face image is different from the original face image.

The second generating unit is configured to generate the first correction loss function according to the difference information, where the first correction loss function is Lg=Ladv+Lfcls+Lrec, Ladv=logD(x)+log(1-D(G(x, c)), Lfcls=-logD(G(x,c)), Lrec=||xG(G(x,c),c')||, G represents the face image generation network, G(x,c) Represents the input image x, the corresponding face image when the generated angle is c, D(x) represents the probability distribution of the domain label calculated by D on the real image, D represents the discriminant network function, c'=-c.

The correction unit is configured to correct the parameters in the pre-trained face image generation network according to the first correction loss function.

In this embodiment, if the second target face image is different from the original face image, or the similarity between the second target face image and the original face image is less than the preset threshold, it means that the reliability of the face image generation network is not To meet the usage requirements, the first correction loss function is Lg, and the face image generation network G is corrected by the first correction loss function. G(x,c) means to generate a face image with an angle of c, G(G(x,c),c') means that the generated image with an angle of c is used as input again to generate the person corresponding to its original angle c' Face image, c'=-c. Lfcls is used to optimize the generation network G, that is, the face image generated by the generation network G can be classified into the target angle domain c by the discrimination network D. Lrec is to ensure that the original face image x is transformed in the angle domain c, and then transformed back to the face image corresponding to the original angle domain c', which is consistent with the original face image.

Further, the device for generating a face image includes:

The judging module is used to judge whether the angle classification corresponding to each first target face image is correct.

The correction module is used to correct the discrimination network through the second correction loss function if not, where the second correction loss function Ld=-Ladv+Lfcls, Ladv=logD(x)+log(1-D(G(x, c)), Lfcls=-logD(G(x,c)), G(x,c) represents the input image x, when the generated angle is c, the corresponding face image, D(x) represents the calculation of the real image by D The probability distribution of the obtained domain label, D represents the discriminant network function.

Further, the device for generating a face image includes:

The obtaining module is used to obtain face images of a specified number of users, where the face images of the same user include at least two angle images corresponding to the specified angles.

The mapping module is used to map the specified angle corresponding to the angle image to the vector value of the specified length.

The combination module is used to combine the angle image and the vector value of the specified length corresponding to the specified angle into a training sample.

The second input module is used to input training samples into the neural network for training to obtain a classification model that can classify the image domain corresponding to the specified angle.

As a module, it is used to use the classification model as a pre-trained face image generation network.

Further, the combination module includes:

The second acquiring unit is configured to acquire the first pixel value corresponding to the facial organ in the first angle image, where the first angle image is included in all angle images.

The third acquiring unit is used to acquire the standard face image corresponding to the first designated angle.

The correction unit is used to correct the first pixel value according to the standard pixel value corresponding to each facial organ in the standard face image to obtain the corrected first pixel value.

The combination unit is used to combine the second angle image carrying the corrected first pixel value and the first vector value corresponding to the first angle into a corrected training sample.

In this embodiment, the training samples are corrected by positioning to improve the accuracy of the training samples. In this embodiment, the face image is centered by cropping, and then the pixel values of the feature points corresponding to the facial organs are corrected. For example, if the standard face image is 256*256, the standard pixel values of the eyes are 58*58, and 158*58. By correcting the first pixel value corresponding to the facial organ in the current user’s first angle image, the correction is aligned to The standard pixel value of the standard face image corresponding to the first specified angle, and the feature points corresponding to other face organs are correspondingly corrected in the above process to reduce the degree of sample difference, the impact on the training accuracy of the face image generation network, and improve the generation network Simultaneously output the accuracy of face images corresponding to multiple angles.

Referring to FIG. 3, an embodiment of the present application also provides a computer device. The computer device may be a server, and its internal structure may be as shown in FIG. 3. The computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor designed by the computer is used to provide calculation and control capabilities. The memory of the computer device includes a readable storage medium and an internal memory. The readable storage medium stores an operating system, computer readable instructions, and a database. The above-mentioned readable storage medium includes a non-volatile readable storage medium and a volatile readable storage medium. The memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium. The database of the computer equipment is used to store the generated face images and other data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer-readable instruction is executed, it executes the process of the above-mentioned method embodiment. Those skilled in the art can understand that the structure shown in FIG. 3 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.

An embodiment of the present application also provides a computer-readable storage medium on which computer-readable instructions are stored. When the computer-readable instructions are executed, the processes of the foregoing method embodiments are executed. The above-mentioned readable storage medium includes non-volatile readable storage medium and volatile readable storage medium. The above are only the preferred embodiments of this application, and do not limit the scope of this application. Any equivalent structure or equivalent process transformation made using the content of the specification and drawings of this application, or directly or indirectly applied to other related The technical field is equally included in the scope of patent protection of this application.

Claims

A method for generating a face image, characterized in that it includes:

Input the angle vectors of the original face image and the target face image into the pre-trained face image generation network at the same time, wherein the angle vector of the target face image includes at least one;

Through the pre-trained face image generation network, the original face image is converted into a first target face image corresponding to a first angle vector according to a preset conversion method, wherein the first angle vector is included in all In the angle vector corresponding to the target face image, the first target face image is included in all the target face images;

Evaluating whether the first angle vector matches the first target face image through a preset discrimination network;

If it matches, output the first target face image.
The method for generating a face image according to claim 1, wherein the pre-trained face image generation network converts the original face image into a first angle vector according to a preset conversion method The corresponding steps of the first target face image include:

Cropping out an image of a human face head region from the original human face image;

Locating the feature points corresponding to the designated organs of the face according to the feature point positioning model, wherein the designated organs include at least one type;

Acquiring coordinate data of feature points corresponding to each of the designated organs;

Mapping each of the coordinate data to the feature point template corresponding to the first angle vector in a one-to-one correspondence to obtain the face area corresponding to the first angle vector;

The first target face image is formed according to the face area corresponding to the first angle vector.
The method for generating a face image according to claim 1, wherein the step of evaluating whether the first angle vector matches the first target face image through a preset discrimination network comprises:

Inputting the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network;

Generating a second target face image according to the inverse angle of the first angle vector and the first target face image;

Judging whether the second target face image is the same as the original face image;

If they are the same, it is determined that the first angle vector matches the first target face image.
The method for generating a face image according to claim 3, wherein after the step of determining whether the second target face image is the same as the original face image, the method comprises:

If the second target face image is not the same as the original face image, determine the difference information between the second target face image and the original face image;

A first correction loss function is generated according to the difference information, where the first correction loss function is Lg=Ladv+Lfcls+Lrec, Ladv=logD(x)+log(1-D(G(x,c)) , Lfcls=-logD(G(x,c)), Lrec=||xG(G(x,c),c')||, G represents the face image generation network, G(x,c) represents the input image x, the corresponding face image when the generated angle is c, D(x) represents the probability distribution of the domain label calculated by D on the real image, D represents the discriminant network function, c'=-c;

The parameters in the pre-trained face image generation network are corrected according to the first correction loss function.
The method for generating a face image according to claim 1, wherein after the step of outputting the first target face image, the method comprises:

Judging whether the angle classification corresponding to each of the first target face images is correct;

If not, the discrimination network is corrected by the second correction loss function, where the second correction loss function Ld=-Ladv+Lfcls, Ladv=logD(x)+log(1-D(G(x,c)) , Lfcls=-logD(G(x,c)), G(x,c) represents the input image x, when the generated angle is c, the corresponding face image, D(x) represents the domain calculated by D on the real image Label probability distribution, D represents the discriminant network function.
The method for generating a face image according to claim 1, wherein before the step of simultaneously inputting the angle vectors of the original face image and the target face image into the pre-trained face image generation network, include:

Acquiring face images of a specified number of users, where the face images of the same user include at least two angle images corresponding to specified angles;

The specified angle corresponding to the angle image is mapped to a vector value of a specified length;

Combining the angle image and the specified length vector value corresponding to the specified angle into a training sample;

Inputting the training samples into a neural network for training to obtain a classification model that can classify the image domain corresponding to the specified angle;

The classification model is used as the pre-trained face image generation network.
The method for generating a face image according to claim 6, wherein the step of combining the angle image and the specified length vector value corresponding to the specified angle into a training sample comprises:

Acquiring a first pixel value corresponding to a human face organ in a first angle image, where the first angle image is included in all the angle images;

Acquiring a standard face image corresponding to the first specified angle;

Correcting the first pixel value according to the standard pixel value corresponding to each facial organ in the standard face image to obtain the corrected first pixel value;

The second angle image carrying the corrected first pixel value and the first vector value corresponding to the first angle are combined into a corrected training sample.
A device for generating a face image, characterized in that it comprises:

The first input module is configured to simultaneously input the angle vectors of the original face image and the target face image into the pre-trained face image generation network, wherein the angle vector of the target face image includes at least one;

The conversion module is configured to use the pre-trained face image generation network to convert the original face image into a first target face image corresponding to a first angle vector according to a preset conversion method, wherein the first The angle vector is included in the angle vectors corresponding to all the target face images, and the first target face image is included in all the target face images;

An evaluation module, configured to evaluate whether the first angle vector matches the first target face image through a preset discrimination network;

The output module is configured to output the first target face image if it matches.
The apparatus for generating a face image according to claim 8, wherein the conversion module comprises:

A cropping unit, configured to crop an image of the head region of the face from the original face image;

The positioning unit is configured to locate the characteristic points corresponding to the designated organs of the face according to the characteristic point positioning model, wherein the designated organs include at least one type;

The first acquiring unit is configured to acquire the coordinate data of the characteristic points corresponding to each of the designated organs;

A mapping unit, configured to map each of the coordinate data to the feature point template corresponding to the first angle vector in a one-to-one correspondence to obtain the face area corresponding to the first angle vector;

The forming unit is configured to form the first target face image according to the face area corresponding to the first angle vector.
The apparatus for generating a face image according to claim 8, wherein the evaluation module comprises:

An input unit, configured to input the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network;

A first generating unit, configured to generate a second target face image according to the inverse angle of the first angle vector and the first target face image;

A judging unit, configured to judge whether the second target face image is the same as the original face image;

The determining unit is configured to determine that the first angle vector matches the first target face image if they are the same.
The apparatus for generating a face image according to claim 10, wherein the evaluation module comprises:

A determining unit, configured to determine difference information between the second target face image and the original face image if the second target face image is different from the original face image;

The second generating unit is configured to generate a first correction loss function according to the difference information, where the first correction loss function is Lg=Ladv+Lfcls+Lrec, Ladv=logD(x)+log(1-D( G(x,c)), Lfcls=-logD(G(x,c)), Lrec=||xG(G(x,c),c')||, G represents the face image generation network, G( x,c) represents the input image x, the corresponding face image when the generated angle is c, D(x) represents the probability distribution of the domain label calculated by D on the real image, D represents the discriminant network function, c'=-c;

The correction unit is configured to correct the parameters in the pre-trained face image generation network according to the first correction loss function.
The apparatus for generating a face image according to claim 8, characterized in that it comprises:

A judging module for judging whether the angle classification corresponding to each of the first target face images is correct;

The correction module is used to correct the discrimination network through a second correction loss function, if not, where the second correction loss function Ld=-Ladv+Lfcls, Ladv=logD(x)+log(1-D(G( x,c)), Lfcls=-logD(G(x,c)), G(x,c) represents the input image x, when the generated angle is c, the corresponding face image, D(x) represents the real The probability distribution of the domain label obtained by the image calculation, D represents the discriminant network function.
The apparatus for generating a face image according to claim 8, characterized in that it comprises:

An acquiring module for acquiring face images of a specified number of users, wherein the face images of the same user include at least two angle images corresponding to the specified angles;

A mapping module for mapping a specified angle corresponding to the angle image to a vector value of a specified length;

A combination module, configured to combine the angle image and the specified length vector value corresponding to the specified angle into a training sample;

The second input module is configured to input the training samples into the neural network for training, and obtain a classification model that can classify the image domain corresponding to the specified angle;

As a module, it is used to use the classification model as the pre-trained face image generation network.
The apparatus for generating a face image according to claim 13, wherein the combination module comprises:

The second acquiring unit is configured to acquire the first pixel value corresponding to the facial organ in the first angle image, wherein the first angle image is included in all the angle images;

The third acquiring unit is configured to acquire a standard face image corresponding to the first designated angle;

The correction unit is configured to correct the first pixel value according to the standard pixel value corresponding to each facial organ in the standard face image to obtain the corrected first pixel value;

The combination unit is configured to combine the second angle image carrying the corrected first pixel value and the first vector value corresponding to the first angle into a corrected training sample.
A computer device includes a memory and a processor, the memory stores a computer program, and is characterized in that, when the processor executes the computer program, a method for generating a face image is realized, and the method for generating a face image includes:

Input the angle vectors of the original face image and the target face image into the pre-trained face image generation network at the same time, wherein the angle vector of the target face image includes at least one;

Through the pre-trained face image generation network, the original face image is converted into a first target face image corresponding to a first angle vector according to a preset conversion method, wherein the first angle vector is included in all In the angle vector corresponding to the target face image, the first target face image is included in all the target face images;

Evaluating whether the first angle vector matches the first target face image through a preset discrimination network;

If it matches, output the first target face image.
The computer device according to claim 15, wherein the pre-trained face image generation network converts the original face image into a first angle vector corresponding to a first angle vector according to a preset conversion method. The steps of the target face image include:

Cropping out an image of a human face head region from the original human face image;

Locating the feature points corresponding to the designated organs of the face according to the feature point positioning model, wherein the designated organs include at least one type;

Acquiring coordinate data of feature points corresponding to each of the designated organs;

Mapping each of the coordinate data to the feature point template corresponding to the first angle vector in a one-to-one correspondence to obtain the face area corresponding to the first angle vector;

The first target face image is formed according to the face area corresponding to the first angle vector.
15. The computer device according to claim 15, wherein the step of evaluating whether the first angle vector matches the first target face image through a preset discrimination network comprises:

Inputting the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network;

Generating a second target face image according to the inverse angle of the first angle vector and the first target face image;

Judging whether the second target face image is the same as the original face image;

If they are the same, it is determined that the first angle vector matches the first target face image.
A computer-readable storage medium with a computer program stored thereon, characterized in that, when the computer program is executed by a processor, a method for generating a face image is realized, and the method for generating a face image includes:

Input the angle vectors of the original face image and the target face image into the pre-trained face image generation network at the same time, wherein the angle vector of the target face image includes at least one;

Through the pre-trained face image generation network, the original face image is converted into a first target face image corresponding to a first angle vector according to a preset conversion method, wherein the first angle vector is included in all In the angle vector corresponding to the target face image, the first target face image is included in all the target face images;

Evaluating whether the first angle vector matches the first target face image through a preset discrimination network;

If it matches, output the first target face image.
The computer-readable storage medium of claim 18, wherein the pre-trained face image generation network converts the original face image into a first angle vector corresponding to a preset conversion method The steps of the first target face image include:

Cropping out an image of a human face head region from the original human face image;

Locating the feature points corresponding to the designated organs of the face according to the feature point positioning model, wherein the designated organs include at least one type;

Acquiring coordinate data of feature points corresponding to each of the designated organs;

Mapping each of the coordinate data to the feature point template corresponding to the first angle vector in a one-to-one correspondence to obtain the face area corresponding to the first angle vector;

The first target face image is formed according to the face area corresponding to the first angle vector.
18. The computer-readable storage medium according to claim 18, wherein the step of evaluating whether the first angle vector matches the first target face image through a preset discrimination network comprises:

Inputting the inverse angle of the first angle vector and the first target face image as input to the pre-trained face image generation network;

Generating a second target face image according to the inverse angle of the first angle vector and the first target face image;

Judging whether the second target face image is the same as the original face image;

If they are the same, it is determined that the first angle vector matches the first target face image.