Three-dimensional face construction method, system, device and medium based on 2D key points
Through the method based on 2D key points, the amount of input data and calculation is reduced, and the problem of slow reconstruction of traditional 3D three-dimensional faces is solved, achieving faster three-dimensional face construction and improving user experience.
Patent Information
- Application Number
- CN202310453617.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-25
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-04-25
AI Technical Summary
The traditional 3D three-dimensional face reconstruction method has a slow reconstruction speed due to the large amount of input data and the large amount of calculation, which affects the user experience.
Using a 2D key point-based method, by obtaining the 2D key point information of the real face image and pairing it with the first mapping relationship, obtaining the second mapping relationship, and only inputting the 2D key point information for three-dimensional face construction, reducing the input data volume and Calculate the amount of calculation, avoiding the gradual iteration process.
It improves the speed and user experience of three-dimensional face construction, and achieves faster three-dimensional face reconstruction.
Smart Images

Figure CN117593447B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of virtual social interaction, and in particular to a method, system, device and medium for constructing a three-dimensional face based on 2D key points. Background Art
[0002] In recent years, with the development and popularization of social networks, the demand for video chat in unfamiliar social scenarios has increased. In this scenario, in order to protect user privacy, anonymous video chat has become a very common choice. In order to ensure the user's anonymous video chat experience, a three-dimensional face is usually used to replace the user's real face. The three-dimensional face is a 3D face that can reflect the user's facial movements and expressions. The traditional 3D face reconstruction method is: take each pixel point of the entire face image as input, use a step-by-step iterative optimization method, and achieve 3D face reconstruction by fitting the expression basis, shape basis, texture basis and pose coefficients. In the traditional method, the input data is a large number of pixels of the face image, which has the problems of large input data volume and large calculation amount, resulting in slow 3D face reconstruction. In addition, the speed of reconstructing 3D faces using step-by-step iterative algorithms is also relatively slow, affecting the user experience. Summary of the invention
[0003] The purpose of the present application is to provide a 3D face construction method, system, device and medium based on 2D key points, which pairs the 2D key point information of multiple real face images with a first mapping relationship to obtain a second mapping relationship. When constructing a 3D face, only the 2D key point information needs to be input, and there is no need to input a large number of pixel points corresponding to the real face image, thereby reducing the amount of input data and calculation, and there is no need to perform a step-by-step iterative process, thereby improving the speed of 3D face construction and user experience.
[0004] In order to solve the above technical problems, the present application provides a method for constructing a three-dimensional face based on 2D key points, which is characterized by comprising:
[0005] Acquire multiple real face images and a first mapping relationship, where the first mapping relationship is a correspondence relationship between each of the real face images and a three-dimensional reconstruction coefficient corresponding to the real face image;
[0006] Obtaining 2D key point information corresponding to each of the real face images according to the multiple real face images and the preset 2D key point model;
[0007] Obtaining a second mapping relationship according to each of the 2D key point information and the first mapping relationship, and training a preset 3D reconstruction coefficient prediction model according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each of the 2D key point information and the 3D reconstruction coefficient corresponding to itself;
[0008] According to the target 2D key point information and the preset 3D reconstruction coefficient prediction model, the target three-dimensional reconstruction coefficient corresponding to the target 2D key point is obtained, and the three-dimensional human face is constructed according to the target three-dimensional reconstruction coefficient.
[0009] Preferably, the 2D key point information includes at least key point information of cheek edges and / or key point information of facial features edges of the real face image.
[0010] Preferably, the 2D key point information includes coordinate information of the 2D key point on the real face image.
[0011] Preferably, the coordinate information includes a horizontal coordinate and a vertical coordinate.
[0012] Preferably, when the 2D key point information includes the horizontal coordinate and the vertical coordinate of the 2D key point on the real face image, obtaining the target three-dimensional reconstruction coefficient corresponding to the target 2D key point according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship includes:
[0013] Convert the target 2D key point information from a two-dimensional array to a one-dimensional array;
[0014] Inputting the target 2D key point information of the one-dimensional array into the preset 3D reconstruction coefficient prediction model;
[0015] A target three-dimensional reconstruction coefficient corresponding to the target 2D key point is obtained according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship.
[0016] Preferably, the preset 3D reconstruction coefficient prediction model includes at least a linear mapping layer and a nonlinear activation layer.
[0017] Preferably, the preset 3D reconstruction coefficient prediction model also includes a batch normalization (BN) layer.
[0018] To solve the above technical problems, the present application also provides a 3D face construction system based on 2D key points, including:
[0019] An acquisition training data unit is used to acquire a plurality of real face images and a first mapping relationship, wherein the first mapping relationship is a correspondence relationship between each of the real face images and a three-dimensional reconstruction coefficient corresponding to the real face image;
[0020] A 2D key point determination unit, configured to obtain 2D key point information corresponding to each of the real face images according to the plurality of real face images and a preset 2D key point model;
[0021] a mapping relationship determination unit, configured to obtain a second mapping relationship according to each of the 2D key point information and the first mapping relationship, and to train a preset 3D reconstruction coefficient prediction model according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each of the 2D key point information and the 3D reconstruction coefficient corresponding to itself;
[0022] The three-dimensional face construction unit is used to obtain target three-dimensional reconstruction coefficients corresponding to the target 2D key points according to the target 2D key point information and the preset 3D reconstruction coefficient prediction model, and construct the three-dimensional face according to the target three-dimensional reconstruction coefficients.
[0023] In order to solve the above technical problems, the present application also provides a 3D face construction device based on 2D key points, comprising:
[0024] Memory for storing computer programs;
[0025] The processor is used to implement the steps of the method for constructing a three-dimensional face based on 2D key points as described above when storing a computer program.
[0026] To solve the above technical problems, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the three-dimensional face construction method based on 2D key points as described above are implemented.
[0027] The present application provides a method, system, device and medium for constructing a three-dimensional face, which relates to the field of virtual social interaction. The scheme obtains corresponding 2D key point information based on multiple real face images and a preset 2D key point model; obtains a second mapping relationship including each 2D key point information and each three-dimensional reconstruction coefficient based on each 2D key point information and a first mapping relationship, and obtains a preset 3D reconstruction coefficient prediction model based on the second mapping relationship; obtains the corresponding target three-dimensional reconstruction coefficient based on the target 2D key point information and the preset 3D reconstruction coefficient prediction model, and constructs a three-dimensional face based on the target three-dimensional reconstruction coefficient. In the present application, the 2D key point information of multiple real face images obtained is paired with the first mapping relationship to obtain a second mapping relationship. When constructing a three-dimensional face, only the 2D key point information needs to be input, and there is no need to input a large number of pixel points corresponding to the real face image, which reduces the amount of input data and the amount of calculation, and there is no need to perform a step-by-step iterative process, which improves the speed of three-dimensional face construction and user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the prior art and the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0029] Figure 1 A flowchart of a method for constructing a three-dimensional face based on 2D key points provided in this application;
[0030] Figure 2 A schematic diagram of a network model in the prior art;
[0031] Figure 3 A schematic diagram of a preset 3D reconstruction coefficient prediction model provided in this application;
[0032] Figure 4 A structural block diagram of a 3D face construction system based on 2D key points provided in this application;
[0033] Figure 5 This is a structural block diagram of a 3D face construction system based on 2D key points provided in this application. DETAILED DESCRIPTION
[0034] The core of the present application is to provide a 3D face construction method, system, device and medium based on 2D key points, which pairs the 2D key point information of multiple real face images with a first mapping relationship to obtain a second mapping relationship. When constructing a 3D face, only the 2D key point information needs to be input, and there is no need to input a large number of pixel points corresponding to the real face image, which reduces the amount of input data and calculation, and does not need to perform a step-by-step iterative process, thereby improving the speed of 3D face construction and user experience.
[0035] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0036] Specifically, the three-dimensional face described in this application may be a 3D headgear or expression that covers the user's real face to ensure the user's privacy.
[0037] At present, the common way to create a three-dimensional face is to use each pixel in a real face image as input data, and use the convolutional neural network to iterate step by step to reconstruct the three-dimensional face by fitting the expression basis, shape basis, texture basis and posture coefficient. Alternatively, a preset model is trained through the correspondence between each real face image and the corresponding three-dimensional reconstruction coefficient, and then the face image is input into the preset model to obtain the three-dimensional reconstruction coefficient. However, both methods require the input of a large number of pixels and contain a lot of redundant information such as background, so the speed of constructing a three-dimensional face is slow.
[0038] Please refer to Figure 1 , Figure 1 A schematic diagram of a process of constructing a 3D face based on 2D key points provided in this application, the method comprising:
[0039] S11: Acquire multiple real face images and a first mapping relationship, where the first mapping relationship is a correspondence relationship between each real face image and a 3D reconstruction coefficient corresponding to the real face image;
[0040] The main idea in this application is to use the 2D key point information of the face instead of each pixel in the real face image to predict the 3D reconstruction coefficients of the face. In this way, the input of the network model is converted from a large number of pixels to a small number of 2D key points, which can greatly improve the speed of 3D reconstruction of the face.
[0041] In the above idea, it is necessary to obtain the three-dimensional reconstruction coefficient from the 2D key point information, so it is necessary to obtain the correspondence between the real face image-2D key point information-3D reconstruction coefficient, and the currently available data is the correspondence between the real face image and the three-dimensional reconstruction coefficient, that is, the first mapping relationship. Therefore, in this embodiment, it is necessary to first obtain the correspondence between the 2D key point information and the three-dimensional reconstruction coefficient, that is, the second mapping relationship, before the 2D key point information can be used as input to obtain the three-dimensional reconstruction coefficient to construct the corresponding three-dimensional face.
[0042] Therefore, in the present application, it is necessary to first obtain original data, which includes multiple real face images and corresponding first mapping relationships, so as to obtain the second mapping relationship based on the original data later.
[0043] In a specific embodiment, when original data is pre-stored, such as the original data that may be data that already exists in the prior art, the original data is stored in a preset storage module, and when it is needed, the original data can be directly read from the preset storage module.
[0044] In another specific embodiment, if the original data is not pre-stored, the real face image and the corresponding first mapping relationship can be trained by the model in the prior art, and stored in the preset storage module, and then when the original data needs to be called, it can be called from the preset storage module.
[0045] S12: obtaining 2D key point information corresponding to each real face image according to multiple real face images and a preset 2D key point model;
[0046] Specifically, after the real face image is obtained as described above, the 2D key point information corresponding to the real face image is extracted according to each real face image and the preset 2D key point model.
[0047] Among them, the preset 2D key point model is the same as the model in the prior art, and this application will not go into details here.
[0048] As a preferred embodiment, the 2D key point information includes at least key point information of cheek edges and / or key point information of facial features edges of a real face image.
[0049] For example, at least key point information corresponding to the edges of eyebrows, eyes, nose, and mouth, as well as key point information of a plurality of facial contours are included. S13: A second mapping relationship is obtained according to each 2D key point information and the first mapping relationship, wherein the second mapping relationship is a correspondence relationship between each 2D key point information and a 3D reconstruction coefficient corresponding to the 2D key point information;
[0050] S13: obtaining a second mapping relationship according to each 2D key point information and the first mapping relationship, and obtaining a preset 3D reconstruction coefficient prediction model through training according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each 2D key point information and a 3D reconstruction coefficient corresponding to itself;
[0051] Specifically, after obtaining the 2D key point information corresponding to each real face image as mentioned above, since there is a corresponding relationship between each real face image and the three-dimensional reconstruction coefficient corresponding to itself, the corresponding relationship between each real face image, the 2D key point information corresponding to itself, and the three-dimensional reconstruction coefficient corresponding to itself can be obtained, so that the corresponding relationship between each 2D key point information and the three-dimensional reconstruction coefficient corresponding to itself can be obtained, that is, the second mapping relationship, and then learning and training are performed based on the second mapping relationship to obtain a preset 3D reconstruction coefficient prediction model, so as to facilitate the subsequent acquisition of the three-dimensional reconstruction coefficient based on the preset 3D reconstruction coefficient prediction model and the 2D key point information.
[0052] In a specific embodiment, the correspondence between each 2D key point information and the corresponding 3D reconstruction coefficient is a one-to-one correspondence. Each real face image is a real face image of the same user or different users.
[0053] S14: Obtain target three-dimensional reconstruction coefficients corresponding to the target 2D key points according to the target 2D key point information and a preset 3D reconstruction coefficient prediction model, and construct a three-dimensional face according to the target three-dimensional reconstruction coefficients.
[0054] After obtaining the second mapping relationship as mentioned above, after receiving the target 2D key point information, the target 2D key point information is input into the preset 3D reconstruction coefficient prediction model, and the target three-dimensional reconstruction coefficient corresponding to the target 2D key point information can be obtained. Based on this target three-dimensional reconstruction coefficient, a three-dimensional face corresponding to the target 2D key point information can be constructed.
[0055] The method for constructing a three-dimensional face in the present application relates to the field of virtual social interaction. The scheme obtains corresponding 2D key point information based on multiple real face images and a preset 2D key point model; obtains a second mapping relationship including each 2D key point information and each three-dimensional reconstruction coefficient based on each 2D key point information and a first mapping relationship, and obtains a preset 3D reconstruction coefficient prediction model based on the second mapping relationship; obtains the corresponding target three-dimensional reconstruction coefficient based on the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship, and constructs a three-dimensional face based on the target three-dimensional reconstruction coefficient. In the present application, the 2D key point information of multiple real face images obtained is paired with the first mapping relationship to obtain the second mapping relationship. When constructing a three-dimensional face, only the 2D key point information needs to be input, and there is no need to input a large number of pixel points corresponding to the real face image, which reduces the amount of input data and the amount of calculation, and there is no need to perform a step-by-step iterative process, which improves the speed of three-dimensional face construction and user experience.
[0056] Based on the above embodiment: As a preferred embodiment, the 2D key point information includes coordinate information of the 2D key point on the real face image.
[0057] Specifically, the 2D key point information obtained in the present application can be the coordinate information of the 2D key point on the real face image. As a preferred embodiment, the coordinate information includes a horizontal coordinate and a vertical coordinate. Obtaining the 2D key point information is to obtain a two-dimensional array including a horizontal coordinate and a vertical coordinate. In a specific embodiment, this two-dimensional array can be directly input into a preset 3D reconstruction coefficient prediction model to obtain the corresponding three-dimensional reconstruction coefficient.
[0058] In a specific embodiment, the 2D key point information includes at least 106 key point information, such as 106 key points. The present application does not limit the number, position and order of the 2D key point information. As long as the key points of the cheek edges and facial features edges are included, they can be used.
[0059] As a preferred embodiment, when the 2D key point information includes the horizontal coordinate and the vertical coordinate of the 2D key point on the real face image, the target three-dimensional reconstruction coefficient corresponding to the target 2D key point is obtained according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship, including:
[0060] Convert the target 2D key point information from a two-dimensional array to a one-dimensional array;
[0061] Input the target 2D key point information of the one-dimensional array into the preset 3D reconstruction coefficient prediction model;
[0062] The target three-dimensional reconstruction coefficients corresponding to the target 2D key points are obtained according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship.
[0063] Considering that when the two-dimensional array including the horizontal coordinate and the vertical coordinate is directly input into the preset 3D reconstruction coefficient prediction model, a neural network layer needs to be set to realize the function of dimensionality reduction of the two-dimensional array, and the structure of the network model is relatively complex.
[0064] In the present application, after obtaining a two-dimensional array including horizontal and vertical coordinates, the two-dimensional array is directly converted into a one-dimensional array, and the one-dimensional array is directly input into a preset 3D reconstruction coefficient prediction model. At this time, there is no need to set a dimensionality reduction neural network layer in the preset 3D reconstruction coefficient prediction model, which simplifies the structure of the preset 3D reconstruction coefficient prediction model.
[0065] As a preferred embodiment, the preset 3D reconstruction coefficient prediction model at least includes a linear mapping layer and a nonlinear activation layer.
[0066] Please refer to Figure 2 , Figure 2 It is a schematic diagram of a network model in the prior art. In the prior art, when using pixel points as input to obtain 3D reconstruction coefficients through a network model, the network model includes at least n Conv2D (2D convolution layer) + Batch Norm (BN, batch normalization) + Relu (non-linear activation layer) modules, and finally + Pooling (pooling layer) + Linear (linear mapping layer) (this network model is only a basic structure example in the prior art, and there may be some other extensions and variations).
[0067] Please refer to Figure 3 , Figure 3A schematic diagram of a preset 3D reconstruction coefficient prediction model provided for the present application. Specifically, when the 2D key point information is used as input, the 2D array of the 2D key point information (such as 106*2) is converted into a 1-dimensional array (such as 312*1), and the Linear operator can well support the operation of the 1-dimensional array. Therefore, this embodiment replaces the Conv2D layer in the original network with Linear. Since the input has become 1-dimensional data, no dimensionality reduction operation is required. The Pooling layer is removed in this embodiment. The preset 3D reconstruction coefficient prediction model in this application at least includes n linear mapping layers Linear + nonlinear activation layers Relu, which simplifies the structure of the preset 3D reconstruction coefficient prediction model. As a preferred embodiment, the preset 3D reconstruction coefficient prediction model also includes a batch normalization BN layer. In order to optimize the training results, the preset 3D reconstruction coefficient prediction model also includes at least a batch normalization BN layer.
[0068] The loss function used in the training process can be consistent with other models that use real face images as input. Let the 2D key point information corresponding to the real face image input to the network be x, the preset 3D reconstruction coefficient prediction model be N, the output result of the preset 3D reconstruction coefficient prediction model be p, and the corresponding annotation result true value be p_gt, then the training loss function is:
[0069] Loss = ||w·(p_gt-p)|| 2 +||V(p_gt)-V(p)|| 2 ;
[0070] Among them, w represents different weights, V represents the process of using three-dimensional reconstruction parameters, and p represents the three-dimensional reconstruction process, which is the same as the three-dimensional reconstruction method in the prior art and will not be described in detail here.
[0071] Please refer to Figure 4 , Figure 4 A structural block diagram of a 3D face construction system based on 2D key points provided in this application, the system includes:
[0072] The training data acquisition unit 21 is used to acquire a plurality of real face images and a first mapping relationship, where the first mapping relationship is a correspondence relationship between each real face image and a 3D reconstruction coefficient corresponding to the real face image;
[0073] A 2D key point determination unit 22, configured to obtain 2D key point information corresponding to each real face image according to a plurality of real face images and a preset 2D key point model;
[0074] A mapping relationship determination unit 23 is used to obtain a second mapping relationship according to each 2D key point information and the first mapping relationship, and to train a preset 3D reconstruction coefficient prediction model according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each 2D key point information and the 3D reconstruction coefficient corresponding to itself;
[0075] The 3D face construction unit 24 is used to obtain the target 3D reconstruction coefficients corresponding to the target 2D key points according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship, and construct the 3D face according to the target 3D reconstruction coefficients. For the introduction of the 3D face construction system based on 2D key points, please refer to the above embodiment, and this application will not repeat it here.
[0076] Please refer to Figure 5 , Figure 5 A structural block diagram of a 3D face construction system based on 2D key points provided in this application, the device includes:
[0077] A memory 31, used for storing computer programs;
[0078] The processor 32 is used to implement the steps of the above-mentioned method for constructing a three-dimensional face based on 2D key points when storing a computer program. For the introduction of the device for constructing a three-dimensional face based on 2D key points, please refer to the above-mentioned embodiment, and this application will not repeat it here.
[0079] To solve the above technical problems, 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 for constructing a three-dimensional face based on 2D key points are implemented. For the introduction of the computer-readable storage medium, please refer to the above-mentioned embodiment, and the present application will not repeat it here.
[0080] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.
[0081] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for constructing a three-dimensional face based on 2D key points, characterized in that: include: Acquire multiple real face images and a first mapping relationship, where the first mapping relationship is a correspondence relationship between each of the real face images and a three-dimensional reconstruction coefficient corresponding to the real face image; Obtaining 2D key point information corresponding to each of the real face images according to the multiple real face images and the preset 2D key point model; Obtaining a second mapping relationship according to each of the 2D key point information and the first mapping relationship, and training a preset 3D reconstruction coefficient prediction model according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each of the 2D key point information and the 3D reconstruction coefficient corresponding to itself; Obtaining target three-dimensional reconstruction coefficients corresponding to the target 2D key points according to the target 2D key point information and the preset 3D reconstruction coefficient prediction model, and constructing the three-dimensional human face according to the target three-dimensional reconstruction coefficients; When the 2D key point information includes the horizontal coordinate and the vertical coordinate of the 2D key point on the real face image, obtaining the target three-dimensional reconstruction coefficient corresponding to the target 2D key point according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship includes: Convert the target 2D key point information from a two-dimensional array to a one-dimensional array; Inputting the target 2D key point information of the one-dimensional array into the preset 3D reconstruction coefficient prediction model; A target three-dimensional reconstruction coefficient corresponding to the target 2D key point is obtained according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship.
2. The method for constructing a three-dimensional face based on 2D key points according to claim 1, characterized in that: The 2D key point information includes at least key point information of cheek edges and / or key point information of facial features edges of the real face image.
3. The method for constructing a three-dimensional face based on 2D key points as claimed in claim 1, characterized in that: The 2D key point information includes coordinate information of the 2D key point on the real face image.
4. The method for constructing a three-dimensional face based on 2D key points as claimed in claim 3, characterized in that: The coordinate information includes a horizontal coordinate and a vertical coordinate.
5. The method for constructing a three-dimensional face based on 2D key points according to any one of claims 1 to 4, characterized in that: The preset 3D reconstruction coefficient prediction model includes at least a linear mapping layer and a nonlinear activation layer.
6. The method for constructing a three-dimensional face based on 2D key points as claimed in claim 5, characterized in that: The preset 3D reconstruction coefficient prediction model also includes a batch normalization (BN) layer.
7. A 3D face construction system based on 2D key points, characterized in that: include: An acquisition training data unit is used to acquire a plurality of real face images and a first mapping relationship, wherein the first mapping relationship is a correspondence relationship between each of the real face images and a three-dimensional reconstruction coefficient corresponding to the real face image; A 2D key point determination unit, configured to obtain 2D key point information corresponding to each of the real face images according to the plurality of real face images and a preset 2D key point model; a mapping relationship determination unit, configured to obtain a second mapping relationship according to each of the 2D key point information and the first mapping relationship, and to train a preset 3D reconstruction coefficient prediction model according to the second mapping relationship, wherein the second mapping relationship is a correspondence relationship between each of the 2D key point information and the 3D reconstruction coefficient corresponding to itself; A three-dimensional face construction unit, used to obtain target three-dimensional reconstruction coefficients corresponding to the target 2D key points according to the target 2D key point information and the preset 3D reconstruction coefficient prediction model, and construct the three-dimensional face according to the target three-dimensional reconstruction coefficients; When the 2D key point information includes the horizontal coordinate and the vertical coordinate of the 2D key point on the real face image, the 3D face construction unit is specifically used to: The target 2D key point information is converted from a two-dimensional array to a one-dimensional array; the target 2D key point information of the one-dimensional array is input into the preset 3D reconstruction coefficient prediction model; the target three-dimensional reconstruction coefficient corresponding to the target 2D key point is obtained according to the target 2D key point information, the preset 3D reconstruction coefficient prediction model and the second mapping relationship, and the three-dimensional face is constructed according to the target three-dimensional reconstruction coefficient.
8. A 3D face construction device based on 2D key points, characterized in that: include: Memory for storing computer programs; A processor, used for implementing the steps of the method for constructing a three-dimensional face based on 2D key points as described in any one of claims 1 to 6 when storing a computer program.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the method for constructing a three-dimensional face based on 2D key points according to any one of claims 1 to 6 are implemented.
