Head reconstruction method based on single photo

Abstract

The invention discloses a head reconstruction method based on a single photo, and the method employs a high-low model mesh fusion algorithm to guarantee the facial details, achieves the model lightweight, and guarantees that the model can adapt to a mobile device and meet the real-time rendering demands. Through combination of key point detection and a PnP algorithm, accurate projection and extraction of facial textures are realized, and the fidelity of texture details is effectively improved. Besides, a skin color matching method based on a Lab color space is introduced, the problem of chromatic aberration between the face area and the prefabricated chartlet is solved, natural edge transition and vivid visual effect are ensured, and therefore a high-quality head three-dimensional model meeting the requirements of multiple platforms and multiple scenes is generated.

Description

Technical Field

[0001] The present application belongs to the technical field of three-dimensional digital humans, and specifically relates to a head reconstruction method based on a single photo. Background Art

[0002] 3D reconstruction technology is an important direction in computer vision and has been widely used in the fields of virtual reality, augmented reality, medical imaging, games, and digital humans. Traditional reconstruction methods mostly rely on multi-view images or depth sensors for data collection. Although these solutions have excellent accuracy, they are limited by high hardware requirements, complex shooting conditions, and high costs, making them difficult to popularize in application scenarios for ordinary users.

[0003] In recent years, 3D reconstruction based on a single photo has become a research hotspot. This technology uses deep learning models and facial statistical models (such as 3D Morphable Model) to parse and reconstruct a single face image. Although these methods have made some progress in predicting facial structure, due to the lack of multi-view information, the model still lacks details, especially in the restoration of complex facial features and expressions. Moreover, most existing 3D reconstruction technologies rely on multi-view image acquisition or depth sensors to generate accurate models, which not only requires expensive equipment, but also places high demands on the shooting environment, making it difficult for ordinary users to use conveniently.

[0004] In addition, facial texture reconstruction is an important part of achieving realism, but existing solutions still face challenges in seamlessly blending facial textures and processing skin color consistency. Color differences are prone to occur between pre-made textures and textures generated from user photos, affecting the overall visual effect. At the same time, insufficient precision in processing occluded areas (such as hair) and edges can also cause unnatural transitions in textures during use. Summary of the invention

[0005] The purpose of this application is to provide a head reconstruction method based on a single photo to solve the problems of high complexity and poor fusion visual effect of traditional three-dimensional reconstruction of human face technology.

[0006] According to the above purpose, the first aspect of the present application provides a head reconstruction method based on a single photo, the method comprising:

[0007] Get the user's face image;

[0008] Generate a high-modulus mesh model M_h from the acquired face image based on the FFHQ-UV algorithm, where M_h=(V_h, F_h), V_h is a vertex set, and F_h is a face set;

[0009] The generated high-poly mesh model M_h is merged with the lightweight low-poly mesh M_l to generate a lightweight low-poly mesh M_l';

[0010] After adjusting the eyeball position on the low-poly mesh M_1', a final head model mesh M_2 is obtained;

[0011] Detect multiple facial key points in the original face image and align them with multiple pre-labeled key points in the head model mesh M_2, project the head model mesh M_2 into the two-dimensional face image, and extract the facial texture map T_face;

[0012] The occluded area is excluded by the face segmentation algorithm to generate a mask area, and a gradient transition area image mask T_transition is generated at the edge of the mask area;

[0013] Generate a skin color reference map C_ref, match the prefabricated map T_pre with the skin color reference map C_ref, and generate an adjusted map T_pre';

[0014] Use the generated mask T_transition containing the transition area to weightedly fuse the texture map T_face of the user's face with the adjusted pre-made map T_pre' to form the spliced ​​head model map T_final;

[0015] Load and render the generated model (mesh M_2, T_final) to form a 3D reconstructed model rendering.

[0016] Further, the generated high-poly mesh model M_h is merged with the lightweight low-poly mesh M_l to generate the lightweight low-poly mesh M_l', which includes: for each vertex v_l on the low-poly mesh M_l, finding the corresponding three nearest vertices v_h1, v_h2, and v_h3 on the high-poly mesh M_h, calculating the weight coefficient, and using linear interpolation to generate the lightweight low-poly mesh M_l', as shown below:

[0017] v_l=w_1*v_h1+w_2*v_h2+w_3*v_h3, where w_1, w_2, w_3 are the weights of each nearest vertex, and w_1+w_2+w_3=1.

[0018] Furthermore, adjusting the eye position on the low-poly mesh M_l' includes: finding the vertices of the eye area of ​​the low-poly mesh M_l' according to the mesh index, and optimizing and adjusting the eye position, and the optimization process uses a curve fitting method: V_smooth = α*V_original+(1-α)*V_neighbors, V_smooth represents the new position of the vertex after smoothing, V_original represents the original position of the vertex, α is the smoothing coefficient, and V_neighbors is a set of adjacent vertices.

[0019] Furthermore, projecting the head model mesh M_2 onto a two-dimensional face image includes: detecting 68 facial key points P_img={p_1, p_2, ..., p_68} in the original face image, and aligning them with the 68 pre-annotated key points P_mesh in mesh M_2, solving the projection matrix M_p of the model using the PnP (Perspective-n-Point) algorithm, and mapping the three-dimensional key points to the plane where the two-dimensional face image is located:

[0020] M_p=K*[R|t];

[0021] Where K is the camera intrinsic parameter matrix, and [R|t] is the rotation and translation matrix.

[0022] Furthermore, matching the prefabricated map T_pre and the skin color reference map C_ref includes: converting the prefabricated map T_pre and the skin color reference map C_ref into Lab color space, and adjusting the value range of the L channel so that the prefabricated map T_pre matches the skin color reference map C_ref, and the adjusted map is represented as T_pre'. At the same time, the brightness of the prefabricated eye map is adjusted to obtain the adjusted eye maps T_lefteye and T_righteye.

[0023] Furthermore, after acquiring the facial image of the user, a face detection algorithm and a posture estimation algorithm are used to evaluate the quality of the image. If no face is detected, or the detected posture angle does not meet the preset posture angle, the user is prompted to retake the photo.

[0024] Furthermore, the texture map T_face of the user's face is weightedly fused with the adjusted prefabricated map T_pre' using the following formula:

[0025] T_final=T_transition*T_face+(1-T_transition)*T_pre', T_final is the head model image after splicing.

[0026] According to a second aspect of the present application, a computer storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the head reconstruction method as described in the first aspect is executed.

[0027] The present application discloses a head reconstruction method based on a single photo, which uses a high-low modulus mesh fusion algorithm to achieve lightweight model while ensuring facial details, ensuring that the model can adapt to mobile devices and real-time rendering requirements. Through the combination of key point detection and PnP algorithm, accurate projection and extraction of facial textures are achieved, effectively improving the fidelity of texture details. In addition, a skin color matching method based on Lab color space is introduced to solve the color difference problem between the facial area and the prefabricated map, ensuring natural edge transitions and realistic visual effects, thereby generating a high-quality head three-dimensional model that meets the needs of multiple platforms and multiple scenarios. Finally, efficient and lightweight head reconstruction is achieved through a single photo, and the model is guaranteed to be rich in details and natural in texture to meet the needs of mobile applications and real-time rendering. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of a high-modulus mesh model M_h in an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of a lightweight mesh model M_1' in an embodiment of the present invention;

[0030] Figure 3 is a schematic diagram of a head mesh M_2 in an embodiment of the present invention;

[0031] Figure 4 is a schematic diagram of an extracted facial texture map T_face in an embodiment of the present invention;

[0032] FIG5( a) is a schematic diagram of a mask area mask_face in an embodiment of the present invention;

[0033] FIG5( b ) is a schematic diagram of a transition region T_transition generated in an embodiment of the present invention;

[0034] Figure 6 is a schematic diagram of a skin color reference image C_ref generated in an embodiment of the present invention;

[0035] FIG. 7( a ) is a schematic diagram of an eyeball map T_lefteye in an embodiment of the present invention;

[0036] FIG7( b ) is a schematic diagram of an eyeball map T_righteye in an embodiment of the present invention;

[0037] Figure 8is a schematic diagram of a facial map T_final obtained by fusion in an embodiment of the present invention;

[0038] Fig. 9 This is a schematic diagram of the effect of model reconstruction displayed after rendering in an embodiment of the present invention;

[0039] Fig.10 Schematic diagram of the flow of a head reconstruction method based on a single photo in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0041] It should be noted that the diagrams provided in this embodiment are only used to illustrate the basic concept of the present invention. The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Any modification of the structure, change in the proportion relationship or adjustment of the size should still fall within the scope of the technical content disclosed by the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0042] The directions or positional relationships indicated by the words "front", "rear", "left", "right", "middle", "longitudinal", "lateral", "horizontal", "inside", "outside" and the like in this specification are based on the directions or positional relationships shown in the drawings and are only for the convenience of simplifying the description. They do not indicate or imply that the devices or elements referred to must have a specific direction, be constructed and operate in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0043] The embodiment of the present application provides a head reconstruction method based on a single photo, which achieves lightweight model while ensuring facial details by adopting a high-low modulus mesh fusion algorithm, ensuring that the model can adapt to mobile devices and real-time rendering requirements; through the combination of key point detection and PnP algorithm, accurate projection and extraction of facial textures are achieved, effectively improving the fidelity of texture details. In addition, a skin color matching method based on Lab color space is introduced to solve the color difference problem between the facial area and the prefabricated map, ensuring natural edge transition and realistic visual effects, thereby generating a high-quality head three-dimensional model that meets the needs of multiple platforms and multiple scenarios.

[0044] Reference Figures 1 to 10As shown, a head reconstruction method based on a single photo disclosed in an embodiment of the present invention specifically includes the following process:

[0045] 1. Obtain the user's facial image;

[0046] Specifically, a high-quality image I containing the user's face can be taken by a mobile phone, ordinary camera, industrial camera or other visual device to ensure that the image contains a complete face and has uniform lighting to avoid the influence of strong light or shadow on texture information. When collecting images, the user maintains a frontal face angle and ensures that the face is not blocked to reduce errors in reconstruction. In addition, the front or rear camera can be selected to shoot according to application requirements to obtain higher resolution and details.

[0047] In order to ensure the accuracy of the collected user face image I, the face detection algorithm D_f(I) and the posture estimation algorithm P(I) can be used to evaluate the image quality. If no face is detected, that is, D_f(I) = 0, or the detected posture angle does not meet the requirements (such as too large or face deflection), the system will prompt the user to retake the photo. The judgment of the posture angle is based on the following formula:

[0048] P(I)=(yaw,pitch,roll), where|yaw|<θ_y,|pitch|<θ_p;

[0049] If any posture angle exceeds the threshold θ_y or θ_p, the system determines that the shooting has failed and returns a prompt to reshoot.

[0050] Based on the FFHQ-UV algorithm, the acquired face image is used to generate a high-modulus mesh model M_h, such as Figure 1 As shown, M_h = (V_h, F_h), V_h is a vertex set, and F_h is a face set.

[0051] Specifically, the FFHQ-UV method is used to generate a high-modulus mesh model M_h = (V_h, F_h) for image I, where V_h is a vertex set and F_h is a facet set. The algorithm optimizes the distribution and structure of mesh vertices so that the generated model can show the geometric details of the face, such as the nose, eye corners and lip contours. In order to improve accuracy, the FFHQ-UV algorithm combines Laplace smoothing technology to reduce noise. Among them, FFHQ-UV is a public project of Tencent Lab and will not be repeated here.

[0052] Second, the generated high-poly mesh model M_h is fused with the lightweight low-poly mesh M_l to generate a lightweight low-poly mesh M_l', such as Figure 2 As shown;

[0053] Specifically, after generating the high-poly mesh M_h, the system fuses it with the lightweight low-poly meshM_l obtained by the wrapping algorithm. In order to establish a correspondence between mesh models of different resolutions, it can ensure that the consistency of features and shapes is retained when the model resolution is changed. The implementation process is: for each vertex v_l on the low-poly mesh M_l, find the corresponding three nearest vertices v_h1, v_h2, and v_h3 on the high-poly mesh M_h, calculate the Euclidean distance with all points on the high-poly mesh, sort and take the smallest three points, which are the nearest three points, and then calculate the weight coefficients w_1, w_2, and w_3, and use linear interpolation to generate the lightweight mesh M_l'. The formula is as follows: v_l=w_1*v_h1+w_2*v_h2+w_3*v_h3,where w_1+w_2+w_3=1.

[0054] The calculation method of the weight coefficient is as follows:

[0055] Assume that a low-poly vertex v finds three nearest vertices p1, p2, and p3 on the high-polygon, and its distances to these three vertices are d1, d2, and d3 respectively. Then the weight of each vertex can be calculated according to the following formula:

[0056]

[0057] Among them, ω i represents the weight of the i-th high-poly vertex to the low-poly vertex, d i is the distance between the i-th high-poly vertex and the low-poly vertex, is a normalization factor so that the sum of all weights is 1.

[0058] The weights generated by the nearest neighbor and weighted average methods can capture and effectively transmit the local geometric information of the high-poly model. In this way, in different high-poly models, as long as the structure and features are similar, these weights can be interpolated at the corresponding positions, so that the low-poly model maintains consistent shape and details. This method retains the key details of the high-poly model while reducing the number of vertices, making the model suitable for use in mobile devices and real-time rendering scenarios, and these weights can be directly used to generate low-poly models from other high-poly models.

[0059] 4. After adjusting the eyeball position on the low-poly mesh M_1', the final head model meshM_2 is obtained, as shown in FIG. Figure 3 shown.

[0060] Specifically, due to the complex geometric features of the eyeball and eye socket areas, the system needs to optimize the mesh vertices in these areas. Since the FLAME head model needs to be reconstructed, the common facial mesh model FLAME usually provides predefined area indexes. For example, the vertex index of the FLAME eyeball is 3931~5022. Through these indexes, the vertex V_eyes of the eyeball position can be found and its position can be adjusted through a smoothing algorithm. By adjusting the position of each vertex to the average of its neighboring vertices, the drastic changes in the vertex position are reduced. The basic idea is to gradually let the vertex approach the average position of its neighbors, smooth the local network, prevent the occurrence of model penetration, and obtain the final head mesh M_2. The optimization process uses a curve fitting method:

[0061] V_smooth=α*V_original+(1-α)*V_neighbors,

[0062] Among them, V_original represents the original position of the vertex. In each smoothing iteration, the new smoothing position can be calculated based on the initial original vertex position. V_smooth represents the new position of the vertex after smoothing, that is, the new position to which the current vertex is to be updated after this smoothing iteration. α is the smoothing coefficient, and V_neighbors is the set of neighboring vertices, which ensures that the eye area is visually real and natural.

[0063] in, Figure 1 to Figure 3 In the model shown, Figure 1 The model is a reconstructed original model using FFHQ-UV technology. The model structure is relatively complex and cannot be seamlessly integrated with the mainstream body model. To solve this problem, and considering the future application requirements in full body reconstruction and efficient driving scenarios, we simplified and optimized the model and finally transformed it into a Figure 2 The lightweight model shown here is not only easier to integrate with mainstream body models, but also has higher practicality and flexibility. Figure 3 Yes Figure 2 Optimization of local positions, especially translation and smoothing of eye positions, makes the eye positions more natural.

[0064] 5. Detect multiple facial key points in the original face image and align them with multiple pre-labeled key points in the head model mesh M_2, project the head model mesh M_2 into the two-dimensional face image, and extract the facial texture map T_face.

[0065] Specifically, the system detects 68 facial key points P_img={p_1,p_2,...,p_68} in image I through an algorithm, and aligns them with the 68 pre-annotated key points P_mesh in the mesh M_2 model. The PnP (Perspective-n-Point) algorithm is used to solve the projection matrix M_p of the model and map the three-dimensional key points to the two-dimensional image plane:

[0066] M_p=K*[R|t];

[0067] Among them, K is the camera intrinsic matrix, and [R|t] is the rotation and translation matrix. This process ensures the accurate registration between the model and the image. After projecting mesh M_2 onto the two-dimensional image, the facial texture map T_face can be extracted, as shown in Figure 2 shown.

[0068] Among them, 68 key points are selected in this embodiment, which basically cover the main geometric features of the face and can effectively describe and locate the important parts of the face. This setting is derived from the classic dlib library and 300-W face data set, and has been widely used in various face detection and recognition tasks. The alignment here is to accurately map the three-dimensional key points to the two-dimensional image plane, ensuring that the projection result of the three-dimensional model in the image space is as consistent as possible with the position of the real face key points in the image, thereby achieving spatial alignment of the image key points and the model key points.

[0069] 6. Exclude the occluded area through the face segmentation algorithm, generate the mask area, and generate the gradient transition area image mask T_transition at the edge of the mask area.

[0070] Specifically, after extracting the facial texture map T_face, the system excludes the occluded area (such as hair) through the face segmentation algorithm and generates a mask area mask_face, as shown in Figure 5(a), and generates a gradient transition area mask T_transition at the edge of the mask area, as shown in Figure 5(b), to ensure a smooth transition of the map during fusion.

[0071] 7. Generate a skin color reference map C_ref, match the prefabricated map T_pre with the skin color reference map C_ref, and generate an adjusted map T_pre';

[0072] Specifically, a skin color reference image C_ref can be generated by using the mask image, such as Figure 4 As shown, this picture is used to adjust the pre-made texture skin color and facial skin color to match, because the final generated texture face is the user's, and the other parts are pre-made, so a fusion operation is required.

[0073] 8. Use the generated mask T_transition containing the transition area to weightedly fuse the texture map T_face of the user's face with the adjusted prefabricated map T_pre' to form a spliced ​​head model map T_final;

[0074] Specifically, the prefabricated texture T_pre and the skin color reference image C_ref are converted to the Lab color space, and the value range of the L channel is adjusted to match the prefabricated texture with the reference image. The adjusted texture is represented as T_pre'. In addition, the brightness of the prefabricated eye texture is adjusted to obtain the adjusted eye textures T_lefteye and T_righteye, so that the eye area is displayed more realistically on the face.

[0075] Afterwards, the mask T_transition containing the transition area generated in step 6 is used to weightedly fuse the texture map T_face of the user's face with the adjusted prefabricated map T_pre' to ensure seamless connection and natural transition between the face and the prefabricated part, and finally the spliced ​​head model map T_final is obtained, as shown in Figure 5.

[0076] T_final=T_transition*T_face+(1-T_transition)*T_pre'

[0077] This method effectively solves the problem of splicing different materials and colors, making the generated head model more natural and realistic visually.

[0078] In addition, in this step, you can also export the optimized model and some maps, such as T_final, mesh M_2, T_lefteye, T_righteye, into multiple formats, such as .obj, .fbx, .glb, etc., to ensure the compatibility of the model in different platforms and engines.

[0079] 9. Load and render the generated model (mesh M_2, T_final) to form a 3D reconstructed model rendering.

[0080] For example, Meshlab can be used to load and render the generated model (M_2, T_final) to display the model's geometry and texture effects, so that users can verify the results of the reconstruction. For example, the model data mesh M_2 can be directly exported as an .obj file by calling a general library function, and the texture data T_final can be saved as a .png file format. These two files correspond to the model's geometry and texture map, respectively, and can be directly loaded and rendered by mainstream 3D software (such as Blender, Maya, 3ds Max, and Unreal Engine, etc.).

[0081] After the model is reconstructed, the system will automatically load the result to the front-end interface and display the reconstructed 3D model through a rendering scene built with Three.js. This rendering scene includes basic interactive functions such as lighting, rotation, and scaling, so that users can intuitively view the facial details and texture effects of the model. In addition, the system uses an optimization algorithm during rendering to ensure smooth model display even on mobile devices.

[0082] Another embodiment of the present application further provides a computer storage medium on which a computer program is stored, wherein the computer program is executed by a processor to implement the head reconstruction method described in the above embodiment.

[0083] If the above functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0084] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A head reconstruction method based on a single photo, characterized in that: The method comprises: Get the user's face image; Generate a high-modulus mesh model M_h from the acquired face image based on the FFHQ-UV algorithm, where M_h=(V_h, F_h), V_h is a vertex set, and F_h is a face set; The generated high-poly mesh model M_h is merged with the lightweight low-poly mesh M_l to generate a lightweight low-poly mesh M_l'; After adjusting the eyeball position on the low-poly mesh M_1', a final head model mesh M_2 is obtained; Detect multiple facial key points in the original face image and align them with multiple pre-labeled key points in the head model mesh M_2, project the head model mesh M_2 into the two-dimensional face image, and extract the facial texture map T_face; The occluded area is excluded by the face segmentation algorithm to generate a mask area, and a gradient transition area image mask T_transition is generated at the edge of the mask area; Generate a skin color reference map C_ref, match the prefabricated map T_pre with the skin color reference map C_ref, and generate an adjusted map T_pre'; Use the generated mask T_transition containing the transition area to weightedly fuse the texture map T_face of the user's face with the adjusted pre-made map T_pre' to form the spliced ​​head model map T_final; Mesh M_2 and T_final are loaded and rendered to form a 3D reconstructed model image.

2. The head reconstruction method according to claim 1, characterized in that: The generated high-poly mesh model M_h is merged with the lightweight low-poly mesh M_l to generate the lightweight low-poly mesh M_l', including: for each vertex v_l on the low-poly mesh M_l, the corresponding three nearest vertices v_h1, v_h2, and v_h3 are found on the high-poly mesh M_h, the weight coefficient is calculated, and the lightweight low-poly mesh M_l' is generated by linear interpolation, as shown below: v_l=w_1*v_h1+w_2*v_h2+w_3*v_h3, where w_1, w_2, w_3 are the weights of each nearest vertex, and w_1+w_2+w_3=1.

3. The head reconstruction method according to claim 1, characterized in that: Adjusting the eyeball position on the low-poly mesh M_l' includes: finding the vertices of the eyeball area of ​​the low-poly mesh M_l' according to the mesh index, and optimizing the eyeball position, and the optimization process uses a curve fitting method: V_smooth = α*V_original+(1-α)*V_neighbors, V_smooth represents the new position of the vertex after smoothing, V_original represents the original position of the vertex, α is the smoothing coefficient, and V_neighbors is a set of adjacent vertices.

4. The head reconstruction method according to claim 2 or 3, characterized in that: Projecting the head model mesh M_2 onto a two-dimensional face image includes: detecting 68 facial key points P_img={p_1, p_2, ..., p_68} in the original face image, aligning them with the 68 pre-labeled key points P_mesh in mesh M_2, solving the projection matrix M_p of the model using the PnP (Perspective-n-Point) algorithm, and mapping the three-dimensional key points to the plane where the two-dimensional face image is located: M_p=K*[R|t]; Where K is the camera intrinsic parameter matrix, and [R|t] is the rotation and translation matrix.

5. The head reconstruction method according to claim 1, characterized in that: Matching the prefabricated map T_pre and the skin color reference map C_ref includes: converting the prefabricated map T_pre and the skin color reference map C_ref to Lab color space, and adjusting the value range of the L channel so that the prefabricated map T_pre matches the skin color reference map C_ref, and the adjusted map is represented as T_pre'. At the same time, the brightness of the prefabricated eyeball map is adjusted to obtain the adjusted eyeball maps T_lefteye and T_righteye.

6. The head reconstruction method according to claim 5, characterized in that: After acquiring the user's facial image, a face detection algorithm and a posture estimation algorithm are used to evaluate the quality of the image. If no face is detected, or the detected posture angle does not meet the preset posture angle, the user is prompted to retake the photo.

7. The head reconstruction method according to claim 1, characterized in that: The texture map T_face of the user's face is weightedly fused with the adjusted prefabricated map T_pre' using the following formula: T_final=T_transition*T_face+(1-T_transition)*T_pre', T_final is the head model map after splicing.

8. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computer program executes the head reconstruction method according to any one of claims 1 to 7.

Images