A method and device for stylized 3D digital human reconstruction based on attribute-based shape bases

Abstract

The application discloses a kind of based on attribute's shape base's stylized 3D digital person reconstruction method and device, and relates to computer technology field.The method comprises the following steps: obtaining the 2D stylized face figure to be reconstructed;2D stylized face figure is input to the trained stylized digital person reconstruction model;According to 2D stylized face figure and stylized digital person reconstruction model, the prediction coefficient of each shape base is obtained;Linear summation is carried out to the calculation result of each shape base minus the preset neutral face Base and the prediction coefficient of each shape base, and the stylized 3D digital person reconstruction result is obtained.The application can solve the problem of insufficient expression of shape base, and improve the expression of fat and thin and facial features of head model.Especially, the number of capital and other data is added to the reconstruction algorithm, which is a revolutionary shape base creation scheme, and can greatly improve the overall reconstruction similarity, aesthetics and expression of the head.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a stylized 3D digital human reconstruction method and apparatus based on an attribute-based shape basis. Background Technology

[0002] As augmented reality (AR) technology gains popularity, an increasing number of B2B (business-to-business) and B2C (business-to-consumer) digital humans are emerging, including hyper-realistic digital humans, realistic digital humans, and avatars (stylized digital humans). The academic community has established rigorous and objective evaluation metrics for hyper-realistic and realistic digital humans, such as the error between the position of the 3D facial landmark projection and the 2D original real facial landmarks. Furthermore, academic institutions and commercial organizations are acquiring vast amounts of facial data to construct universally recognized 3D facial digital assets. For example, the open-source library BFM provides commonly used academic tools such as neutral face bases, shape bases, expression bases, and base-color textures. Many digital human avatars in academia and business are created based on these universally recognized open-source digital assets.

[0003] However, a common evaluation metric for stylized avatars has yet to emerge, nor has a shared digital asset base for multiple styles been developed. The methods for creating digital assets are still in the exploratory stage. Among the numerous digital assets, the neutral face base and shape base directly determine stylization. These two parts of the digital assets can reconstruct static stylized face models. The type and degree of stylization are determined by the base and shape base, making their creation particularly important. The neutral face base is a stylized face model manually adjusted by experienced modelers, determining the overall reconstructed cartoon image, such as Western Warcraft, Disney cartoons, Japanese anime, etc. The shape base is based on the neutral face base, expanding with over 100 face shape files, reflecting differences in different dimensions of the face. During reconstruction, the face is reconstructed within the upper and lower limits of the expressive power of these shape bases. There are many methods for generating shape bases, each with its advantages and disadvantages. Currently, several benchmark stylized images in academia and business have problems, mainly due to unscientific creation of the shape base.

[0004] In actual use, the coefficients corresponding to these bases are predicted, and then the neutral face base is subtracted from each of the shape bases. The diff result is linearly summed with the predicted coefficients to obtain the reconstruction result.

[0005] There are two main methods for creating stylized shape bases: one is based on PCA (Principal Component Analysis), and the other is based on semantics.

[0006] Basis models created using PCA exhibit good generalization because they synthesize tens of millions of 2D face images into a massive tensor. PCA is then used for dimensionality reduction, prioritizing the top 100+ dimensions of the tensor based on importance, thus using features that can represent a large number of faces as the basis for reconstruction. However, PCA basis models may not fully represent all aspects of a complete face. If insufficient representation is found, it becomes difficult to pinpoint the specific missing dimension. Unlike semantic basis models, PCA face basis models show minimal variation in performance, making it challenging to determine which dimension is causing the problem.

[0007] Semantic-based bases are intuitive, with each dimension representing a visual difference in facial shape, such as wide nostrils, a large nose tip, or an upturned nose. It's straightforward; missing dimensions can be easily added. However, existing semantic bases lack expressive power and cannot adequately depict facial details because their definitions are based on semantic bases used in real-world face reconstruction, which are not entirely applicable to stylized bases. Furthermore, there is the issue of overlapping dimensions. For example, if the first dimension represents a large mouth and the second dimension represents a small mouth, a positive coefficient in the first dimension is equivalent to a negative coefficient in the second, leading to overlap and abnormal reconstruction results with weak generalization. Summary of the Invention

[0008] This invention addresses the problem of insufficient expressive power of shape-based substrates in existing technologies.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0010] On one hand, the present invention provides a stylized 3D digital human reconstruction method based on an attribute-based shape basis, which is implemented by an electronic device and includes:

[0011] S1. Obtain the 2D stylized face image to be reconstructed.

[0012] S2. Input the 2D stylized face image into the trained stylized digital human reconstruction model.

[0013] S3. Based on the 2D stylized face image and the stylized digital human reconstruction model, obtain the prediction coefficients of each shape basis.

[0014] S4. Subtract the calculation result of the preset neutral face base from each shape base and sum linearly with the prediction coefficient of each shape base to obtain the stylized 3D digital human reconstruction result.

[0015] Optionally, the prediction coefficients of the shape basis in S3 include prediction coefficients of continuous shape basis and prediction coefficients of discrete shape basis.

[0016] Optionally, the basis controlled by the prediction coefficients of the continuous shape basis includes:

[0017] Principal Component Analysis (PCA) Basis Based on Real-Life Shape Basis Transfer.

[0018] Slim face base and chubby face base.

[0019] Ear base that is close to the face, large ear base, and small ear base.

[0020] Standard nasal base, wide nasal alar base, and large nasal tip base.

[0021] Short and thick lip base, thin upper lip base and thick lower lip base, moderate thickness of upper and lower lip base, thick upper lip base and thin lower lip base, and thin upper and lower lip base.

[0022] Straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, and inwardly curved thin eyebrow base.

[0023] The outer corner of the eye is level without a drooping base; the outer corner of the eye is high without a drooping base; and the outer corner of the eye is high with a drooping base.

[0024] Optionally, the basis controlled by the prediction coefficients of the discrete shape basis includes:

[0025] Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eyeball pupil color.

[0026] Optionally, the training process for the stylized digital human reconstruction model in S2 includes:

[0027] S21. Obtain training data consisting of multiple 2D stylized face image data and label data consisting of multiple 2D stylized face key point coordinate data.

[0028] S22. Train the stylized digital human reconstruction model based on the training data and label data.

[0029] Optionally, the stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0030] The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis.

[0031] The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients for discrete shape basis.

[0032] Optionally, in S4, the calculated result of subtracting the preset neutral face base from each shape base and linearly summing it with the predicted coefficients of each shape base yields the stylized 3D digital human reconstruction result, including:

[0033] S41. Subtract the preset neutral face base from each shape base and sum linearly with the prediction coefficients of each shape base.

[0034] S42. Substitute the summation result into the differentiable renderer to generate the rendered image.

[0035] S43. By calculating the loss value between the 2D stylized facial key point coordinate data of the 3D projection in the rendered image and the 2D stylized facial key point coordinate data of the label data, the shape of the stylized 3D digital human reconstruction is constrained to obtain the stylized 3D digital human reconstruction result.

[0036] On the other hand, the present invention provides an attribute-based shape-based stylized 3D digital human reconstruction apparatus, which is used to implement an attribute-based shape-based stylized 3D digital human reconstruction method. The apparatus includes:

[0037] The acquisition module is used to acquire the 2D stylized face image to be reconstructed.

[0038] The input module is used to input 2D stylized face images into the trained stylized digital human reconstruction model.

[0039] The prediction coefficient module is used to obtain the prediction coefficients of each shape basis based on the 2D stylized face image and the stylized digital human reconstruction model.

[0040] The output module is used to linearly sum the calculated results of subtracting the preset neutral face base from each shape base and the prediction coefficients of each shape base to obtain the stylized 3D digital human reconstruction results.

[0041] Optionally, the prediction coefficients of the shape basis include prediction coefficients of continuous shape basis and prediction coefficients of discrete shape basis.

[0042] Optionally, the basis controlled by the prediction coefficients of the continuous shape basis includes:

[0043] Principal Component Analysis (PCA) Basis Based on Real-Life Shape Basis Transfer.

[0044] Slim face base and chubby face base.

[0045] Ear base that is close to the face, large ear base, and small ear base.

[0046] Standard nasal base, wide nasal alar base, and large nasal tip base.

[0047] Short and thick lip base, thin upper lip base and thick lower lip base, moderate thickness of upper and lower lip base, thick upper lip base and thin lower lip base, and thin upper and lower lip base.

[0048] Straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, and inwardly curved thin eyebrow base.

[0049] The outer corner of the eye is level without a drooping base; the outer corner of the eye is high without a drooping base; and the outer corner of the eye is high with a drooping base.

[0050] Optionally, the basis controlled by the prediction coefficients of the discrete shape basis includes:

[0051] Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eyeball pupil color.

[0052] Optionally, the input module is further used for:

[0053] S21. Obtain training data consisting of multiple 2D stylized face image data and label data consisting of multiple 2D stylized face key point coordinate data.

[0054] S22. Train the stylized digital human reconstruction model based on the training data and label data.

[0055] Optionally, the stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0056] The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis.

[0057] The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients for discrete shape basis.

[0058] Optionally, the output module is further used for:

[0059] S41. Subtract the preset neutral face base from each shape base and sum linearly with the prediction coefficients of each shape base.

[0060] S42. Substitute the summation result into the differentiable renderer to generate the rendered image.

[0061] S43. By calculating the loss value between the 2D stylized facial key point coordinate data of the 3D projection in the rendered image and the 2D stylized facial key point coordinate data of the label data, the shape of the stylized 3D digital human reconstruction is constrained to obtain the stylized 3D digital human reconstruction result.

[0062] On one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described stylized 3D digital human reconstruction method based on attribute shape basis.

[0063] On the one hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement the above-described stylized 3D digital human reconstruction method based on attribute shape basis.

[0064] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0065] The aforementioned solution proposes a comprehensive shape-based base modeling approach, encompassing the digital face and related facial data such as hair, beard, eyelashes, and glasses, including their location, type, and color. This enhances the accuracy and diversity of 3D stylized digital human reconstruction. Furthermore, by incorporating data beyond the head model into the algorithmic reconstruction process, various types and colors of digital assets can be reconstructed based on the base data, reducing the cost of manually generating data by modelers. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the stylized 3D digital human reconstruction method based on attribute-based shape basis provided in an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the stylized 3D digital human reconstruction method based on attribute-based shape basis provided in an embodiment of the present invention;

[0069] Figure 3 This is a block diagram of a stylized 3D digital human reconstruction device based on an attribute-based shape basis provided in an embodiment of the present invention.

[0070] Figure 4This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0071] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0072] like Figure 1 As shown, this embodiment of the invention provides a stylized 3D digital human reconstruction method based on an attribute-based shape basis, which can be implemented by an electronic device. Figure 1 The flowchart shown is for a stylized 3D digital human reconstruction method based on attribute-based shape basis. The processing flow of this method may include the following steps:

[0073] S1. Obtain the 2D stylized face image to be reconstructed.

[0074] S2. Input the 2D stylized face image into the trained stylized digital human reconstruction model.

[0075] Optionally, the training process for the stylized digital human reconstruction model in S2 includes:

[0076] S21. Obtain training data consisting of multiple 2D stylized face image data and label data consisting of multiple 2D stylized face key point coordinate data.

[0077] S22. Train the stylized digital human reconstruction model based on the training data and label data.

[0078] Optionally, the stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0079] The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis.

[0080] The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients for discrete shape basis.

[0081] In one feasible implementation, a cartoon-style neutral face base is first manually adjusted by a modeler, and then a set of shape bases is created offline in advance. This base includes the differences in various parts of the facial area, as well as the upper and lower limits of the shape of hair, eyelashes, eyebrows, beard, and glasses.

[0082] Furthermore, the algorithm trains the stylized digital human reconstruction model offline. The training data can be tens of thousands of 2D stylized face images, and the label data can be the coordinates of 2D facial key points (normalized: X coordinate divided by image width, Y coordinate divided by image height). The backbone network is ResNet101, which is ultimately connected to two fully connected (FC) layers. One FC layer directly predicts the coefficients of all continuous variables, serving as the BS coefficients for the position and shape of each facial part. The prediction result of the other FC layer is then fed into a softmax layer, whose output serves as the coefficients for predicting the material type, color, and other discrete variables of each facial part.

[0083] S3. Based on the 2D stylized face image and the stylized digital human reconstruction model, obtain the prediction coefficients of each shape basis.

[0084] Optionally, the prediction coefficients of the shape basis in S3 include prediction coefficients of continuous shape basis and prediction coefficients of discrete shape basis.

[0085] Optionally, the basis controlled by the prediction coefficients of the continuous shape basis may include:

[0086] 1-100 Dimensions: PCA base generated based on the migration of a real-life shape base.

[0087] 101-102 dimensions: Slim face base, chubby face base.

[0088] 103-105 dimensions: ear base close to the face, large ear base, small ear base.

[0089] 106-108 dimensions: standard nasal base, wide nasal alar base, large nasal tip base.

[0090] 109-113 dimensions: Short and thick lip base, thin upper lip and thick lower lip base, moderate thickness of both upper and lower lip base, thick upper lip and thin lower lip base, thin upper and lower lip base

[0091] 114-118 dimensions: straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, inwardly curved thin eyebrow base.

[0092] 119-121 dimensions: Level outer corner of the eye without drooping base; High outer corner of the eye without drooping base; High outer corner of the eye with drooping base.

[0093] Optionally, the basis controlled by the prediction coefficients of the discrete shape basis may include:

[0094] Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eyeball pupil color.

[0095] S4. Subtract the calculation result of the preset neutral face base from each shape base and sum linearly with the prediction coefficient of each shape base to obtain the stylized 3D digital human reconstruction result.

[0096] Optionally, step S4 above may include the following steps S41-S43:

[0097] S41. Subtract the preset neutral face base from each shape base and sum linearly with the prediction coefficients of each shape base.

[0098] S42. Substitute the summation result into the differentiable renderer to generate the rendered image.

[0099] S43. By calculating the loss value between the 2D stylized facial key point coordinate data of the 3D projection in the rendered image and the 2D stylized facial key point coordinate data of the label data, the shape of the stylized 3D digital human reconstruction is constrained to obtain the stylized 3D digital human reconstruction result.

[0100] In one feasible implementation, the prediction coefficients are linearly summed with the corresponding basis, substituted into a differentiable renderer to generate a rendered image, and the 3D reconstructed shape is constrained by calculating the loss value between the 2D key point coordinates of the 3D projection and the label key point coordinates in the rendered image.

[0101] Furthermore, in practical use, a styled 2D face image is input, substituted into the trained model, the coefficients corresponding to the shape basis are predicted, and the coordinates of the reconstructed point cloud are calculated by linear summation to complete the stylized reconstruction.

[0102] This invention proposes a comprehensive shape-based base construction scheme based on facial attributes, including a digital face and the location, type, color, etc., of facial-related data such as hair, beard, eyelashes, and glasses, improving the reconstruction accuracy and diversity of 3D stylized digital humans. Furthermore, by incorporating data other than the head model into the algorithmic reconstruction process, various types and colors of digital assets can be reconstructed based on the base data, reducing the cost of manually generating data by modelers.

[0103] like Figure 3 As shown, this embodiment of the invention provides a stylized 3D digital human reconstruction device 300 based on attribute-based shape basis. This device 300 is used to implement a stylized 3D digital human reconstruction method based on attribute-based shape basis. The device 300 includes:

[0104] The acquisition module 310 is used to acquire the 2D stylized face image to be reconstructed.

[0105] Input module 320 is used to input 2D stylized face images into a trained stylized digital human reconstruction model.

[0106] The prediction coefficient module 330 is used to obtain the prediction coefficients of each shape basis based on the 2D stylized face image and the stylized digital human reconstruction model.

[0107] The output module 340 is used to linearly sum the calculation results of subtracting the preset neutral face base from each shape base and the prediction coefficients of each shape base to obtain the stylized 3D digital human reconstruction results.

[0108] Optionally, the prediction coefficients of the shape basis include prediction coefficients of continuous shape basis and prediction coefficients of discrete shape basis.

[0109] Optionally, the basis controlled by the prediction coefficients of the continuous shape basis includes:

[0110] Principal Component Analysis (PCA) Basis Based on Real-Life Shape Basis Transfer.

[0111] Slim face base and chubby face base.

[0112] Ear base that is close to the face, large ear base, and small ear base.

[0113] Standard nasal base, wide nasal alar base, and large nasal tip base.

[0114] Short and thick lip base, thin upper lip base and thick lower lip base, moderate thickness of upper and lower lip base, thick upper lip base and thin lower lip base, and thin upper and lower lip base.

[0115] Straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, and inwardly curved thin eyebrow base.

[0116] The outer corner of the eye is level without a drooping base; the outer corner of the eye is high without a drooping base; and the outer corner of the eye is high with a drooping base.

[0117] Optionally, the basis controlled by the prediction coefficients of the discrete shape basis includes:

[0118] Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eyeball pupil color.

[0119] Optionally, the input module 320 is further used for:

[0120] S21. Obtain training data consisting of multiple 2D stylized face image data and label data consisting of multiple 2D stylized face key point coordinate data.

[0121] S22. Train the stylized digital human reconstruction model based on the training data and label data.

[0122] Optionally, the stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0123] The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis.

[0124] The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients for discrete shape basis.

[0125] Optionally, the output module 340 is further used for:

[0126] S41. Subtract the preset neutral face base from each shape base and sum linearly with the prediction coefficients of each shape base.

[0127] S42. Substitute the summation result into the differentiable renderer to generate the rendered image.

[0128] S43. By calculating the loss value between the 2D stylized facial key point coordinate data of the 3D projection in the rendered image and the 2D stylized facial key point coordinate data of the label data, the shape of the stylized 3D digital human reconstruction is constrained to obtain the stylized 3D digital human reconstruction result.

[0129] This invention proposes a comprehensive shape-based base construction scheme based on facial attributes, including a digital face and the location, type, color, etc., of facial-related data such as hair, beard, eyelashes, and glasses, improving the reconstruction accuracy and diversity of 3D stylized digital humans. Furthermore, by incorporating data other than the head model into the algorithmic reconstruction process, various types and colors of digital assets can be reconstructed based on the base data, reducing the cost of manually generating data by modelers.

[0130] Figure 4 This is a schematic diagram of the structure of an electronic device 400 provided in an embodiment of the present invention. The electronic device 400 can vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 401 and one or more memories 402. The memory 402 stores at least one instruction, which is loaded and executed by the processor 401 to implement the following attribute-based shape-based stylized 4D digital human reconstruction method:

[0131] S1. Obtain the 2D stylized face image to be reconstructed.

[0132] S2. Input the 2D stylized face image into the trained stylized digital human reconstruction model.

[0133] S3. Based on the 2D stylized face image and the stylized digital human reconstruction model, obtain the prediction coefficients of each shape basis.

[0134] S4. Subtract the calculation result of the preset neutral face base from each shape base and sum linearly with the prediction coefficient of each shape base to obtain the stylized 3D digital human reconstruction result.

[0135] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to perform the aforementioned stylized 3D digital human reconstruction method based on attribute-based shape. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0136] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An attribute-based shape-based stylized 3D digital human reconstruction method, characterized in that, The method includes: S1. Obtain the 2D stylized face image to be reconstructed; S2. Input the 2D stylized face image into the trained stylized digital human reconstruction model; S3. Based on the 2D stylized face image and the stylized digital human reconstruction model, obtain the prediction coefficients of each shape basis; S4. Subtract the preset neutral face base from each shape base and sum the results linearly with the prediction coefficients of each shape base to obtain the stylized 3D digital human reconstruction results. The prediction coefficients of the shape basis in S3 include the prediction coefficients of continuous shape basis and the prediction coefficients of discrete shape basis. The basis controlled by the prediction coefficients of the continuous shape basis includes: Principal component analysis (PCA) basis generated from real-person shape basis migration; Slim face base and chubby face base; Ear base that fits close to the face, large ear base, and small ear base; Standard nasal base, wide alar base, and large nasal tip base; Short and thick lip base, thin upper lip base and thick lower lip base, moderate thickness of upper and lower lip base, thick upper lip base and thin lower lip base, and thin upper and lower lip base; Straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, and inwardly curved thin eyebrow base; The outer corner of the eye is level without a drooping base; the outer corner of the eye is high without a drooping base; and the outer corner of the eye is high with a drooping base. The basis controlled by the prediction coefficients of the discrete shape basis includes: Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eyeball pupil color.

2. The method of claim 1, wherein, The training process of the stylized digital human reconstruction model in S2 includes: S21. Obtain training data consisting of multiple 2D stylized face image data and label data consisting of multiple 2D stylized face key point coordinate data. S22. Train the stylized digital human reconstruction model based on the training data and label data.

3. The method according to claim 1, characterized in that, The stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer; The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis; The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients of the discrete shape basis.

4. The method of claim 1, wherein, In step S4, the calculated result of subtracting the preset neutral face base from each shape base and linearly summing it with the prediction coefficients of each shape base yields the stylized 3D digital human reconstruction result, including: S41. The calculation result of subtracting the preset neutral face Base from each shape base is linearly summed with the prediction coefficient of each shape base; S42. Substitute the summation result into the differentiable renderer to generate the rendered image; S43. By calculating the loss value between the 2D stylized facial key point coordinate data of the 3D projection in the rendered image and the 2D stylized facial key point coordinate data of the label data, the shape of the stylized 3D digital human reconstruction is constrained to obtain the stylized 3D digital human reconstruction result.

5. An attribute-based shape-based stylized 3D digital human reconstruction apparatus, characterized in that, The device includes: The acquisition module is used to acquire the 2D stylized face image to be reconstructed; The input module is used to input the 2D stylized face image into the trained stylized digital human reconstruction model; The prediction coefficient module is used to obtain the prediction coefficients of each shape basis based on the 2D stylized face image and the stylized digital human reconstruction model. The output module is used to linearly sum the calculation results of subtracting the preset neutral face base from each shape base and the prediction coefficients of each shape base to obtain the stylized 3D digital human reconstruction results. The prediction coefficients of the shape basis include prediction coefficients for continuous shape basis and prediction coefficients for discrete shape basis; The basis controlled by the prediction coefficients of the continuous shape basis includes: Principal component analysis (PCA) basis generated from real-person shape basis migration; Slim face base and chubby face base; Ear base that fits close to the face, large ear base, and small ear base; Standard nasal base, wide alar base, and large nasal tip base; Short and thick lip base, thin upper lip base and thick lower lip base, moderate thickness of upper and lower lip base, thick upper lip base and thin lower lip base, and thin upper and lower lip base; Straight thick eyebrow base, inwardly curved thick eyebrow base, straight thin eyebrow base, and inwardly curved thin eyebrow base; The outer corner of the eye is level without a drooping base; the outer corner of the eye is high without a drooping base; and the outer corner of the eye is high with a drooping base. The basis controlled by the prediction coefficients of the discrete shape basis includes: Lip color, eyebrow texture, lip texture, eyebrow color, face texture, face skin tone, and eye pupil color.

6. The apparatus of claim 5, wherein, The stylized digital human reconstruction model includes a backbone network ResNet101, a first fully connected layer, a second fully connected layer, and a softmax layer; The first fully connected layer is used to obtain the prediction coefficients of the continuous shape basis; The second fully connected layer connects to the softmax layer and is used to obtain the prediction coefficients of the discrete shape basis.

Images