Single-image three-dimensional head reconstruction method and system

Abstract

This application discloses a single-image three-dimensional head reconstruction method and system. The method includes: obtaining an image to be reconstructed and extracting a head image with the background removed from it; inputting the head image into a trained normal feature extraction model to obtain a normal feature map of the head image; inputting the head image and its normal feature map into a trained image encoder to obtain an image feature map of the head image; inputting voxels within a preset resolution in three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator to obtain the distance value of each voxel in three-dimensional space to the surface of a potential three-dimensional head model; and constructing a three-dimensional head model of the image to be reconstructed using voxels whose distance values ​​are less than a preset threshold. The present invention can reconstruct a high-quality, high-fidelity three-dimensional head model from a single image.

Description

Technical Field

[0001] The present application relates to the fields of computer graphics and computer vision technology, and more specifically, to a single-image three-dimensional head reconstruction method and system. Background Art

[0002] Three-dimensional reconstruction is an important area of ​​focus for researchers in computer vision and graphics. Three-dimensional head reconstruction has broad application prospects in the fields of metaverse, virtual reality, film and television entertainment, medical beauty, etc. There are currently two ways to obtain a three-dimensional head model: (1) Scan the head with precision three-dimensional imaging equipment to obtain a three-dimensional head model; (2) Reconstruct the three-dimensional head model directly from the image using computer vision and graphics technology. The three-dimensional head model obtained by the first method is highly accurate, but the equipment used (such as structured light three-dimensional scanner) is expensive and difficult to operate, making it difficult to popularize; and due to the presence of noise, the obtained three-dimensional model is usually incomplete and has loopholes. The second method directly reconstructs the three-dimensional head model from the photo, which has a high degree of automation, low cost, and broad application prospects. The second method can be further divided into methods based on multi-view images and methods based on single images. Methods based on multi-view generally require the simultaneous acquisition of multiple images and still require certain application conditions; while methods based on single images only require one image, which reduces the application conditions to the greatest extent.

[0003] In recent years, with the rapid development of deep learning, scholars have proposed a series of single-image 3D head reconstruction methods, but most of these methods are based on parametric models and have technical defects. Parametric model-based methods mainly use 3D deformable models (such as the Flame model) to express the head. By estimating limited parameters (such as 99 parameters) from image features, a specific head model can be deformed from the initial model. This type of method is fast and simple, but it cannot express arbitrary facial topological changes and can usually only restore a rough 3D head shape. There are the following disadvantages: (1) The reconstructed head model is not personalized enough and mostly looks like the initial model. (2) It is usually unable to reconstruct details such as facial expression profiles and wrinkles, and is almost powerless for hair. (3) It is difficult to reconstruct across a large age range. For example, a model trained on an adult dataset is not suitable for reconstructing children's heads. Summary of the Invention

[0004] In response to at least one shortcoming or improvement need in the prior art, the present invention provides a single-image three-dimensional head reconstruction method and system, which can reconstruct a high-quality, high-fidelity three-dimensional head model from a single image.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a single-image three-dimensional head reconstruction method is provided, comprising:

[0006] Acquire an image to be reconstructed and extract a head image with background removed therefrom;

[0007] Inputting the head image into a trained normal feature extraction model to obtain a normal feature map of the head image;

[0008] Inputting the head image and its normal feature map into a trained image encoder to obtain an image feature map of the head image;

[0009] Inputting voxels within a preset resolution in the three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator to obtain a distance value from each voxel in the three-dimensional space to the surface of the potential three-dimensional head model;

[0010] The three-dimensional head model of the image to be reconstructed is constructed using voxels whose distance values ​​are less than a preset threshold.

[0011] Furthermore, the normal feature extraction model includes a downsampling block, a feature information enhancement module and an expanded feature map size module connected in sequence, the feature information enhancement module is composed of multiple residual blocks, and the expanded feature map size module is composed of multiple transposed convolution interpolation operations.

[0012] Furthermore, the image encoder is a network composed of convolution blocks and second-order hourglass networks stacked four times.

[0013] Furthermore, the equipotential surface estimator is a 6-layer perceptron, the input layer includes 257 neurons, the middle layer includes 1024, 512, 256, and 128 neurons respectively, and the output layer includes 1 neuron.

[0014] Furthermore, inputting the feature map of the head image into a trained equipotential surface estimator comprises:

[0015] Create a three-dimensional voxel block and convert the three-dimensional coordinates of the voxels in the voxel block into Transformed into coordinates on a two-dimensional feature map through weak projection perspective ;

[0016] The feature map of the head image contains c channels, and the coordinate position retrieved in the feature map of the head image is The image feature values ​​of c channels are connected into a feature vector ;

[0017] The three-dimensional coordinates of the voxels in the voxel block The normalized value of the Z-axis coordinate is used as the depth information value of the voxel ;

[0018] The feature vector and depth information value By inputting it into the trained equipotential surface estimator, the equipotential surface estimator is recorded as , the output of the equipotential surface estimator is recorded as , .

[0019] Furthermore, the calculation formula of the weak projection perspective transformation is:

[0020]

[0021] in, are the camera parameters of the image to be reconstructed, is the scaling factor, is the translation vector, Indicates rounding down.

[0022] Furthermore, a three-dimensional head model sample is obtained, and a head image sample and a normal feature map label are extracted from the three-dimensional head model sample. ;

[0023] The normal feature extraction model is trained using the head image sample and the normal feature map label. The loss function of the normal feature extraction model is for

[0024]

[0025] in, is the normal feature map estimated by the normal feature extraction model, is the weight factor, is the L1 loss, represents the VGG perceptual loss.

[0026] Furthermore, an image feature map of the head image sample is obtained, three-dimensional coordinate points are collected from the three-dimensional head model sample, and distance values ​​from the three-dimensional coordinate points to the surface of the three-dimensional head model sample are calculated as equipotential surface estimation labels of the head image sample;

[0027] The image encoder and the equipotential surface estimator are trained using the head image sample, its image feature map, and the equipotential surface estimation label.

[0028] Furthermore, constructing the three-dimensional head model of the image to be reconstructed by using voxels whose distance values ​​are less than a preset threshold includes:

[0029] For voxels whose distance value is less than the preset threshold, the equipotential surface is extracted through the pipeline matching algorithm to reconstruct the three-dimensional head model, and the three-dimensional head model is smoothed.

[0030] According to a second aspect of the present invention, there is also provided a single-image three-dimensional head reconstruction system, comprising:

[0031] A preprocessing module, used for acquiring an image to be reconstructed and extracting a head image with background removed therefrom;

[0032] a normal feature map extraction module, configured to input the head image into a trained normal feature extraction model to obtain a normal feature map of the head image;

[0033] A reconstruction module is configured to input the head image and its normal feature map into a trained image encoder to obtain an image feature map of the head image, input voxels within a preset resolution in three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator to obtain a distance value from each voxel in three-dimensional space to the surface of a potential three-dimensional head model, and construct a three-dimensional head model of the image to be reconstructed using voxels whose distance values ​​are less than a preset threshold.

[0034] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0035] (1) The present invention uses a method based on normal feature-assisted equipotential surface estimation to perform single-view head modeling. The flexibility of the equipotential surface estimator and the normal feature are used to guide the reconstruction in space. Compared with the traditional head reconstruction method based on parameterized models, the present invention has the following advantages: the reconstructed head model is more personalized and close to the real individual to the greatest extent; it also has a certain reconstruction capability for hair; it can reconstruct a high-fidelity three-dimensional head model with facial details such as large expressions and wrinkles; the invention has a certain reconstruction capability for both children and adults.

[0036] (2) Compared with the method of collecting three-dimensional head data using precise three-dimensional imaging equipment, the present invention only needs to input the portrait image into the computer to complete the reconstruction. It has the advantages of high degree of automation and low cost, and has broad application prospects in the future.

[0037] (3) In the method of directly reconstructing a three-dimensional head model from an image using computer vision and graphics technology, the present invention reduces the application conditions to the greatest extent. Only one picture needs to be input to reconstruct a complete head model, which is more conducive to popularization among the public. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. 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 creative work.

[0039] Figure 1 A schematic diagram of the process of a single-image 3D head reconstruction method provided in an embodiment of the present application;

[0040] Figure 2 A schematic diagram of the network and principle of the single-image 3D head reconstruction method provided in an embodiment of the present application;

[0041] Figure 3 A schematic diagram of a second-order hourglass network substructure of an image encoder provided in an embodiment of the present application;

[0042] Figure 4 Schematic diagram of the smoothed and unsmoothed effects of the three-dimensional head model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0044] The terms "including" and "having," and any variations thereof, in the specification and claims of this application and the accompanying drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or modules is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to the process, method, product, or apparatus.

[0045] like Figure 1 As shown, a single-image three-dimensional head reconstruction method according to an embodiment of the present invention includes the following steps:

[0046] S101, obtaining an image to be reconstructed and extracting a head image from it with the background removed.

[0047] Processing input portrait images that include background , from which the human head mask image is segmented and recorded as , and then use the mask image to limit the area of ​​the input image that the computer operates on:

[0048]

[0049] In one embodiment, an API interface provided by an existing platform can be used to remove the background outside the portrait from the input image, and then a code is used to segment the human head mask image, and the mask image and the original image are equally cropped.

[0050] S102: Input the head image into a trained normal feature extraction model to obtain a normal feature map of the head image.

[0051] The normal features of each pixel are estimated from the input head image to obtain a normal feature map.

[0052] Preferably, a deep neural network is used To estimate the normal features. Figure 2 As shown in the upper middle section, the normal feature extraction model includes a sequentially connected downsampling block, a feature information enhancement module, and a feature map resizing module. The feature information enhancement module is composed of multiple residual blocks, and the feature map resizing module is composed of multiple transposed convolution interpolation operations. Preferably, the residual block is composed of several consecutive convolution operations.

[0053] Specifically, the loss function for training the normal feature extraction model is:

[0054]

[0055] is the normal feature map label of the head image sample The normal feature map estimated by the normal feature extraction model, is the weight factor, is the L1 loss, Represents VGG perceptual loss. Under the constraints of the loss function, the normal feature map obtained by rendering the 3D head model is used as the true label to train the neural network. , set here The clarity of the normal feature map obtained by training with L1 loss alone is much lower than the effect predicted under the joint constraint of adding high-level layer loss. When using it, you only need to pass a head picture into the network to estimate the corresponding normal feature map. Preferably, each pixel in the normal feature map has a certain height value. The normal feature map can intuitively describe the convexity and concavity of the object with very little resource information, and can provide more surface detail information for the image encoder. Figure 2 The upper right corner N in the figure shows the estimated normal feature map.

[0056] S103: Input the head image and its normal feature map into the trained image encoder to obtain an image feature map of the head image.

[0057] Joint input head image and normal feature map After that, the image encoder E is used to extract the image features, and the image feature map is recorded as .

[0058] Preferably, the image encoder E is a deep convolutional neural network, which is a network composed of four stacked convolution blocks and a second-order hourglass network, that is, a network consisting of three convolution operations, a second-order hourglass network, three convolution operations, a second-order hourglass network, three convolution operations, a second-order hourglass network, three convolution operations, and a second-order hourglass network. The structure of a second-order hourglass network is as follows: Figure 3 shown.

[0059] After fusing the normal feature map, the image encoder can obtain a more informative feature vector, enabling the reconstruction of a high-fidelity 3D head model. Specifically, the feature information obtained by the image encoder is both local and global. The local feature facilitates the reconstruction of surface details, while the global feature allows for better prediction even in occluded areas.

[0060] Preferably, the image encoder input is the image of the fusion channel, and output is Feature map ,in =256 is the number of channels of the feature, 、 are the width and height of the feature map respectively, and a total of 4 feature maps are output in sequence.

[0061] S104: Inputting the voxels within a preset resolution in the three-dimensional space and the image feature map of the head image into the trained equipotential surface estimator to obtain the distance value between each voxel in the three-dimensional space and the surface of the potential three-dimensional head model.

[0062] Based on the image features, an equipotential surface estimator is used Estimate each 3D point in 3D space Distance to the surface of the potential 3D head model ; The distance is the same The points form a equipotential surfaces, where The equipotential surface is the surface of the potential three-dimensional head model; the distances from all points in the three-dimensional space to the surface of the potential three-dimensional head model constitute a potential field, such as Figure 2 The medium potential field is shown.

[0063] Preferably, the equipotential surface estimator is a 6-layer perceptron. Specifically, the number of neurons in each layer of the multilayer perceptron is 257, 1024, 512, 256, 128, and 1, respectively. The input data is the superposition of the pixel feature vector and the Z-axis coordinate value of the 3D point, and the output is the predicted distance value between the 3D point and the potential 3D head model. Figure 2 The lower right corner of the figure shows the equipotential surface estimator .

[0064] Preferably, establish a voxel blocks, where , combine the feature vector of each pixel to estimate the distance value of each voxel in the voxel block in the equipotential field, and extract the distance value of voxels, where , and the Marching Cube algorithm is used to reconstruct these voxels into a three-dimensional head model represented by a mesh.

[0065] Preferably, the distance value of a single voxel in the potential field is estimated as follows:

[0066] (1) Coordinate mapping: Create a three-dimensional voxel block and set the three-dimensional coordinates of the selected voxel as ( ), which is transformed to its coordinates on the two-dimensional feature map through weak projection perspective ( ):

[0067]

[0068] in, are the camera parameters estimated from the input image, is the scaling factor, is the translation vector, Indicates rounding down. The camera parameters used in this example are orthographic projection.

[0069] The above processing is performed on each voxel point in the voxel block. At this time, each three-dimensional coordinate has a corresponding two-dimensional projection point on the feature map.

[0070] (2) Feature retrieval: In the image feature map The coordinate position to be retrieved is The image feature values ​​of c channels are connected into a feature vector (denoted as ), size is 256.

[0071] (3) Depth information estimation: Each voxel The normalized value of the Z-axis coordinate of the voxel Depth information value , with a size of 1.

[0072] (4) Distance value estimation: through equipotential surface estimator According to the image feature vector and depth information estimated voxels Distance to the surface of the potential 3D head model:

[0073]

[0074] Preferably, if Figure 2 As shown, the image encoder E and the equipotential surface estimator After obtaining the required training set data items, we can minimize the function To train the image encoder E and the equipotential surface estimator The neural network parameters, The calculation formula is:

[0075]

[0076] Preferably, =0.7; is a set of three-dimensional head mesh sampling points; ; is the label function, ,if is a point outside the sampling grid (i.e., the surface of the three-dimensional model); , if is a point on the sampling grid surface.

[0077] After estimating the distance values ​​from all voxels in the voxel block to the potential three-dimensional head model, the equipotential surfaces are extracted through the pipeline matching algorithm to reconstruct the three-dimensional head model.

[0078] Leveraging the flexibility of equipotential surface estimators’ expressions and normal feature maps to spatially guide equipotential surface estimators It is predicted that this invention can reconstruct a distinctively personalized three-dimensional head model, restore the delicate facial expressions, and has the ability to reconstruct across age stages, that is, it has a certain reconstruction ability for both adults and children.

[0079] Preferably, the equipotential surface is extracted by the pipeline matching algorithm to reconstruct the 3D head model, and the surface of the 3D head model is smoothed. Due to the influence of voxels, there may be small gaps between the point layers in the reconstructed 3D head model. The surface of the reconstructed 3D head model is smoothed by the Laplacian Smooth algorithm to obtain a fine head model, such as Figure 4 As shown, the left side is the 3D head model that has not been smoothly reconstructed, and the right side is the refined head model obtained after surface smoothing.

[0080] The following details the training principles of the normal feature extraction model, image encoder, and equipotential surface estimator.

[0081] (1) Obtain a 3D head model sample and extract the head image sample and normal feature map label from the 3D head model sample ; Use head image samples and normal feature map labels to train the normal feature extraction model. The loss function of the normal feature extraction model is as described above.

[0082] (2) Obtain the image feature map of the head image sample, collect 3D coordinate points from the 3D head model sample, calculate the distance value from the 3D coordinate point to the surface of the 3D head model sample, and use it as the equipotential surface estimation label of the head image sample; use the head image sample and its image feature map and equipotential surface estimation label to train the image encoder and equipotential surface estimator. The loss functions of the image encoder and equipotential surface estimator are as described above.

[0083] In one embodiment, each 3D head mesh is first rendered to obtain a training image and a normal feature map label. Then, the 3D head mesh is spirally sampled to obtain a point set S. The distance values ​​from the 3D points in the point set S to the surface of the 3D head model sample are calculated as the labels corresponding to the training images. Finally, training data items are assembled, each of which includes a head image, a mask image, a normal feature map, a set of head mesh sampling points, and camera parameters.

[0084] The present invention applies implicit representation to single-view Figure 3 In the process of 3D head reconstruction, unlike the existing methods based on parameterized models, the present invention is not limited to expressing only fixed topological structures and can reconstruct a high-fidelity 3D head model.

[0085] A single-image three-dimensional head reconstruction system according to an embodiment of the present invention includes a preprocessing module, a normal feature map extraction module, and a reconstruction module.

[0086] The preprocessing module is used to obtain the image to be reconstructed and extract the head image with the background removed therefrom.

[0087] Preferably, this module primarily segments a head mask from an input portrait image containing a background, limiting the computer's manipulation of the image to the head region. This module encapsulates the background removal API, image cropping, and channel segmentation code to obtain a mask. Specifically, the background of the input portrait image is removed to obtain the mask, and the mask and input image are cropped to a 512x512 resolution.

[0088] The normal feature map extraction module is used to input the head image into the trained normal feature extraction model to obtain the normal feature map of the head image.

[0089] Preferably, the module encapsulates the trained A deep neural network is trained and the trained weight parameter file is retained. After initializing the network model, the above S2 step is executed to use this network to predict the normal feature map corresponding to the input head image, with a size of 3*512*512.

[0090] The reconstruction module is used to input the head image and its normal feature map into a trained image encoder to obtain the image feature map of the head image, input the voxels within a preset resolution in the three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator, obtain the distance value between each voxel in the three-dimensional space and the surface of the potential three-dimensional head model, and use the voxels whose distance values ​​are less than a preset threshold to construct a three-dimensional head model of the image to be reconstructed.

[0091] Preferably, this module encapsulates the image encoder E and the equipotential surface estimator with fixed weight parameters After initializing the neural network, the above steps S3, S4, and S5 are executed in sequence, and the learned image pixel feature information is used as an intermediate medium to guide the equipotential surface estimator. The distance relationship between the 3D points and the underlying 3D model is mapped; finally, a pipeline matching algorithm is used to extract equipotential surfaces and reconstruct the 3D head model. Specifically, in this example, the data obtained from M1 and M2 are input and the reconstructed 3D head model is output. The reconstruction module is used to construct the 3D head model of the image to be reconstructed using voxels whose distance values ​​are less than a preset threshold.

[0092] Preferably, the single-image three-dimensional head reconstruction system further includes a smoothing module for performing surface smoothing processing on the reconstructed three-dimensional head model.

[0093] Preferably, the module encapsulates the Laplacian Smooth algorithm, and the three-dimensional head model reconstructed by M3 can be further smoothed and optimized selectively through step S6, and the final result is retained.

[0094] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of the system or module can be electrical or other forms.

[0096] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of the present embodiment according to actual needs.

[0097] In addition, the functional modules in the various embodiments of the present application may be integrated into a processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module. The above-mentioned integrated modules may be implemented in the form of hardware or software functional modules.

[0098] If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM), random access memory (RAM), mobile hard disk, magnetic disk, or optical disk, etc., various media that can store program code.

[0099] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructing related hardware through a program. The program may be stored in a computer-readable memory, which may include a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0100] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the specification and practicing the disclosure herein, those skilled in the art will easily think of the implementation scheme of the present disclosure. This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

[0101] The technical features of the above embodiments can be combined arbitrarily. 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.

[0102] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A single-image three-dimensional head reconstruction method, characterized in that: include: Acquire an image to be reconstructed and extract a head image with background removed therefrom; Inputting the head image into a trained normal feature extraction model to obtain a normal feature map of the head image; Inputting the head image and its normal feature map into a trained image encoder to obtain an image feature map of the head image; Inputting voxels within a preset resolution in the three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator to obtain a distance value from each voxel in the three-dimensional space to the surface of the potential three-dimensional head model; The three-dimensional head model of the image to be reconstructed is constructed using voxels whose distance values ​​are less than a preset threshold.

2. The single-image 3D head reconstruction method according to claim 1, wherein: The normal feature extraction model includes a downsampling block, a feature information enhancement module and an expanded feature map size module connected in sequence, the feature information enhancement module is composed of multiple residual blocks, and the expanded feature map size module is composed of multiple transposed convolution interpolation operations.

3. The single-image 3D head reconstruction method according to claim 1, wherein: The image encoder is a network composed of four stacked convolution blocks and a second-order hourglass network.

4. The single-image 3D head reconstruction method according to claim 1, wherein: The equipotential surface estimator is a 6-layer perceptron, the input layer includes 257 neurons, the middle layer includes 1024, 512, 256, and 128 neurons respectively, and the output layer includes 1 neuron.

5. The single-image 3D head reconstruction method according to claim 1, wherein: Inputting the feature map of the head image into the trained equipotential surface estimator comprises: Establish a three-dimensional voxel block, and transform the three-dimensional coordinate v of the voxel in the voxel block into the coordinate x on the two-dimensional feature map through weak projection perspective; The feature map of the head image contains c channels. The image feature value at the coordinate position x is retrieved from the feature map of the head image, and the retrieved image feature values ​​of the c channels are connected into a feature vector F x ; The normalized value of the Z-axis coordinate of the three-dimensional coordinate v of the voxel in the voxel block is used as the depth information value D of the voxel v ; The eigenvector F x And the depth information value D v By inputting it into the trained equipotential surface estimator, the equipotential surface estimator is denoted as f, and the output of the equipotential surface estimator is denoted as 6. The single-image 3D head reconstruction method according to claim 5, wherein: The calculation formula of the weak projection perspective transformation is: Among them, (s,t x ,t y ) is the camera parameter of the image to be reconstructed, s is the scaling factor, (t x ,t y ) is the translation vector, Indicates rounding down.

7. The single-image 3D head reconstruction method according to claim 1, wherein: Obtain a 3D head model sample, and extract a head image sample and a normal feature map label from the 3D head model sample The normal feature extraction model is trained using the head image samples and normal feature map labels. The loss function L of the normal feature extraction model is N for Where N is the normal feature map estimated by the normal feature extraction model, w N is the weight factor, L1() represents L1 loss, L VGG () represents VGG perceptual loss.

8. The single-image three-dimensional head reconstruction method according to claim 7, wherein: Obtaining an image feature map of the head image sample, collecting three-dimensional coordinate points from the three-dimensional head model sample, and calculating distance values ​​from the three-dimensional coordinate points to the surface of the three-dimensional head model sample as equipotential surface estimation labels of the head image sample; The image encoder and the equipotential surface estimator are trained using the head image sample, its image feature map, and the equipotential surface estimation label.

9. The single-image 3D head reconstruction method according to claim 1, wherein: The constructing of the three-dimensional head model of the image to be reconstructed by using voxels whose distance values ​​are less than a preset threshold comprises: For voxels whose distance value is less than the preset threshold, the equipotential surface is extracted through the pipeline matching algorithm to reconstruct the three-dimensional head model, and the three-dimensional head model is smoothed.

10. A single-image three-dimensional head reconstruction system, characterized in that: include: A preprocessing module, used for acquiring an image to be reconstructed and extracting a head image with background removed therefrom; a normal feature map extraction module, configured to input the head image into a trained normal feature extraction model to obtain a normal feature map of the head image; A reconstruction module is configured to input the head image and its normal feature map into a trained image encoder to obtain an image feature map of the head image, input voxels within a preset resolution in three-dimensional space and the image feature map of the head image into a trained equipotential surface estimator to obtain a distance value from each voxel in three-dimensional space to the surface of a potential three-dimensional head model, and construct a three-dimensional head model of the image to be reconstructed using voxels whose distance values ​​are less than a preset threshold.

Images