Three-dimensional model reconstruction method and device based on generative adversarial network
By applying the GAN-based three-dimensional model reconstruction method in three-dimensional geological modeling, the problems of insufficient connectivity of spatial structures and large computing resource occupation in traditional methods are solved, and efficient and precise reconstruction of complex geological structures are achieved.
Patent Information
- Application Number
- CN202411499492.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-02-18
AI Technical Summary
Traditional multi-point geological statistics methods have problems such as insufficient spatial structure connectivity and large computing resource occupation in three-dimensional geological modeling, while existing deep learning methods are difficult to effectively reconstruct complex oriented under the limitations of conditional data and generation modes. The underground structure of the opposite sex.
Using a three-dimensional model reconstruction method based on a generative adversarial network (GAN), by obtaining the conditional profile of the original 3-dimensional model, filling and segmenting to obtain training data, building a generative adversarial network including generators and discriminators, and passing binary crossover Entropy loss function and spatial context loss function are trained to generate a more accurate geological model.
By optimizing the structure of the generated adversarial network and extracting data spatial characteristics, this method avoids the problems of discontinuity and unnaturalness of spatial structures in traditional methods, enhances the ability to reconstruct complex geological structures, and improves the reconstruction efficiency and accuracy of reconstruction .
Smart Images
Figure CN119478210A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of three-dimensional geological modeling, and in particular to a three-dimensional model reconstruction method and device based on a generative adversarial network. Background Art
[0002] The deep learning method aims to establish a mapping relationship between the two data sets, the training set and the label set, and use this mapping relationship to complete the test set fitting. The deep learning method usually undergoes multiple trainings to learn the spatial structural features in the training set and establish the mapping relationship between the training set and the label set. Because its network structure has multiple layers of neural networks and a large number of neurons to store the fitting information of the model, the deep learning network can capture more spatial pattern features than the traditional multi-point geostatistical method, enabling it to reconstruct more complex and heterogeneous geological structures. The introduction of a variety of deep neural networks provides an effective technical guarantee for the reconstruction of anisotropic complex geological structures. At present, the deep learning method has become an important branch in the field of automatic modeling, and has been well applied in the fields of geophysical gravity and magnetic inversion, porous media reconstruction, reservoir modeling, and hydrogeological modeling.
[0003] As a widely used method in the field of traditional geological modeling, the three-dimensional modeling method based on multi-point geostatistics can directly obtain the spatial pattern of random variables in the reference model using data templates, and match the data events in the image to be simulated with the spatial pattern in the previously acquired training image, and determine the high-dimensional statistical information of the specific pattern by multiple points. However, whether it is a pattern-based or pixel-based multi-point statistical method, since a pattern matching in its simulation sub-process can only determine the value of a pattern or a pixel, the generated simulation results often have problems such as insufficient spatial structural connectivity or spatial structural distortion. In addition, the three-dimensional modeling method based on multi-point geostatistics is also a typical computationally intensive task, and each model reconstruction process will take up a lot of computing resources. Compared with the multi-point geostatistical modeling method, the deep learning method is good at efficiently extracting spatial patterns in images through multiple neural layers that bear different tasks, and integrating the time and computational overhead of the subsequent modeling process into the training process, providing a new solution for the field of automatic modeling.
[0004] As a deep learning network based on the idea of zero-sum game, generative adversarial network has significant advantages in feature extraction and generation, and therefore has received extensive attention in the fields of geological modeling such as porous media reconstruction, rock parameter prediction, and texture synthesis. However, most of the existing GAN network-based methods realize end-to-end reconstruction of a single two-dimensional image to a three-dimensional model or three-dimensional reconstruction using simple conditional data as constraints. Although the former method can reproduce the spatial pattern provided in the two-dimensional image, it is difficult to add hard data constraints in the reconstruction process; and the latter method has greater limitations in reconstructing underground structures with complex anisotropy due to the limited information that the conditional data can provide. Summary of the invention
[0005] In view of this, the purpose of the present invention is to provide a three-dimensional model reconstruction method and device based on a generative adversarial network, which is used to solve the problems of insufficient connectivity and spatial structure distortion of the reconstructed model of traditional multi-point geostatistical methods and the limitations of conditional data and generation patterns in existing deep learning methods.
[0006] The present invention provides a three-dimensional model reconstruction method based on a generative adversarial network, comprising the steps of:
[0007] S1: obtaining an original three-dimensional model, and extracting multiple conditional sections from the original three-dimensional model;
[0008] S2: Fill and segment each conditional section to obtain a set of training 3D models; segment the original 3D model to obtain a set of labeled 3D models;
[0009] S3: construct a generative adversarial network, train the generative adversarial network by training the 3D model set and the label 3D model set, and obtain a trained generative adversarial network;
[0010] S4: Input the three-dimensional model to be reconstructed into the trained generative adversarial network to obtain the reconstructed three-dimensional model.
[0011] Preferably, step S2 specifically comprises:
[0012] S21: sorting the conditional sections according to their spatial positions in the original three-dimensional model, and filling the blank positions between the conditional sections to generate a filled three-dimensional model of the same size as the original three-dimensional model;
[0013] S22: segmenting the filled three-dimensional model at different scales, and using the obtained filled three-dimensional sub-model sets at different scales as the training three-dimensional model set;
[0014] S23: Segment the original three-dimensional model into different scales, and use the obtained original three-dimensional sub-model sets of different scales as the label three-dimensional model set.
[0015] Preferred:
[0016] Generative adversarial networks include: generator and discriminator;
[0017] The generator includes a first SAGAN network, a first self-attention layer, a U-Net network, a second SAGAN network, and a second self-attention layer connected in sequence, and the first SAGAN network is connected to the second SAGAN network;
[0018] The discriminator includes a first PatchGAN network, a second PatchGAN network, and a linear convolutional network connected in sequence.
[0019] Preferably, step S3 specifically comprises:
[0020] S31: input the training 3D model into the generator to generate a predicted residual 3D model; downsample and upsample the label 3D model in sequence to generate a sampled 3D model, and subtract the label 3D model from the sampled 3D model to obtain a true residual 3D model;
[0021] S32: input the predicted residual three-dimensional model and the true residual three-dimensional model into the discriminator, and calculate and obtain the binary cross entropy loss function and the spatial context loss function;
[0022] S33: Calculate the joint loss function through the binary cross entropy loss function and the spatial context loss function;
[0023] S34: Back-propagate the generative adversarial network through the joint loss function to adjust the parameters of the generative adversarial network;
[0024] S35: Repeat steps S31-S34 until the joint loss function is less than a preset value to obtain a trained generative adversarial network.
[0025] Preferred:
[0026] The calculation formula of the binary cross entropy loss function BCELoss is:
[0027]
[0028] Where i is the number of the training 3D model; N is the total number of training 3D models; y i is the true label of the true residual 3D model corresponding to the training 3D model i, and its value is 0 or 1; p i Represents the probability that the generated prediction residual 3D model is the true label.
[0029] Preferred:
[0030] Spatial context loss function L ctext The calculation formula of (G) is:
[0031]
[0032] Among them, binary_mask(i) is the binary mask matrix, R real is the residual structure of the true residual 3D model, R generated is the residual structure of the predicted residual 3D model, Indicates that the training 3D model i obeys the real data distribution P data Expected sampling.
[0033] Preferred:
[0034] Joint loss function G * The calculation formula is:
[0035] G * =BCELoss+λ1L ctext (G)
[0036] Among them, BCELoss is the binary cross entropy loss function, L ctext (G) is the spatial context loss function, and λ1 is the loss coefficient.
[0037] A storage medium storing instructions and data for implementing the three-dimensional model reconstruction method based on a generative adversarial network.
[0038] A three-dimensional model reconstruction device based on a generative adversarial network includes: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional model reconstruction method based on the generative adversarial network.
[0039] The present invention has the following beneficial effects:
[0040] (1) The generative adversarial network constructed by the present invention adds a self-attention layer to multiple networks of the generator, optimizes the structure of the generative adversarial network, and the generative adversarial network based on the self-attention layer extracts data spatial features through feature maps, avoiding the problem that the generated results contain discontinuous and unnatural spatial structures in the traditional multi-point geological statistics-based three-dimensional reconstruction method.
[0041] (2) By filling and segmenting the conditional profile to obtain training data, the data format of the conditional data is optimized, the amount of spatial structure information carried in the conditional data is increased, the shortcomings of the conditional data constraints in the three-dimensional reconstruction process of the traditional deep learning model are avoided, and the reconstruction capability of underground structures containing complex anisotropy is improved.
[0042] (3) The joint loss function is obtained by calculating the binary cross entropy loss function and the spatial context loss function. The joint loss function is used to guide the generative adversarial network to generate a more accurate geological model from both local and global aspects, thereby improving the efficiency and accuracy of a single reconstruction of the three-dimensional model. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;
[0044] Figure 2 A schematic diagram of the workflow of the present invention and the training process of generating an adversarial network;
[0045] Figure 3 This is a schematic diagram of the three-dimensional reconstruction experiment of binary discrete attributes;
[0046] Figure 4 A schematic diagram showing the comparison of statistical characteristics of simulation results in the high connectivity channel data experiment;
[0047] Figure 5 It is a schematic diagram of the three-dimensional reconstruction experiment of the multi-element discrete attribute stratigraphic structure model;
[0048] Figure 6 It is a schematic diagram for comparing the statistical characteristics of simulation results in the multi-attribute complex geological structure data experiment;
[0049] Figure 7 It is a schematic diagram of the three-dimensional reconstruction experiment of the multi-element discrete attribute channel facies reservoir model;
[0050] Figure 8 It is a schematic diagram for comparing the statistical characteristics of simulation results in the multi-attribute complex geological structure data experiment;
[0051] Fig. 9 This is a structural diagram of the device according to an embodiment of the present invention;
[0052] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0053] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0054] Reference Figure 1 The present invention provides a three-dimensional model reconstruction method based on a generative adversarial network, comprising the steps of:
[0055] S1: obtaining an original three-dimensional model, and extracting multiple conditional sections from the original three-dimensional model;
[0056] Further, the workflow of the present invention is as follows Figure 2As shown, first, multiple sections of three different scales are extracted from the original three-dimensional model and placed according to the original spatial positions of the sections.
[0057] S2: Fill and segment each conditional section to obtain a set of training 3D models; segment the original 3D model to obtain a set of labeled 3D models;
[0058] Furthermore, step S2 is specifically as follows:
[0059] S21: sorting the conditional sections according to their spatial positions in the original three-dimensional model, and filling the blank positions between the conditional sections to generate a filled three-dimensional model of the same size as the original three-dimensional model;
[0060] S22: segmenting the filled three-dimensional model at different scales, and using the obtained filled three-dimensional sub-model sets at different scales as the training three-dimensional model set;
[0061] Specifically, a three-dimensional Gaussian filter up-down sampling method is used to pre-process the filled three-dimensional model, and preferably, a plurality of small three-dimensional blocks of three different scales are generated and used as a set of training three-dimensional models;
[0062] S23: Segment the original three-dimensional model into different scales, and use the obtained original three-dimensional sub-model sets of different scales as the label three-dimensional model set.
[0063] Specifically, the original three-dimensional model is segmented into three small blocks of different scales according to step S22 for use as a set of labeled three-dimensional models. During the segmentation process, the spatial orientation of each small three-dimensional block corresponds to the orientation of the small block in S22.
[0064] S3: construct a generative adversarial network, train the generative adversarial network by training the 3D model set and the label 3D model set, and obtain a trained generative adversarial network;
[0065] Furthermore, the generative adversarial network includes: a generator and a discriminator;
[0066] The generator includes a first SAGAN network, a first self-attention layer, a U-Net network, a second SAGAN network, and a second self-attention layer connected in sequence, and the first SAGAN network is connected to the second SAGAN network;
[0067] The discriminator includes a first PatchGAN network, a second PatchGAN network, and a linear convolutional network connected in sequence.
[0068] Specifically, the SAGAN network is based on transposed convolutions, and a self-attention layer is introduced after the highest convolutional layer of the SAGAN network. This structure can adjust the configuration of the self-attention layer according to the importance of specific geological features, or select an appropriate number of levels according to the complexity of the data, so that the model can be more adapted to specific geological modeling tasks.
[0069] The PatchGAN network used does not make a one-time true or false judgment on the entire image, but divides the image into several small blocks and performs a true or false evaluation on each small block separately, so that the discriminator can focus on the local details of the image and improve its sensitivity to image quality.
[0070] Furthermore, the training process of the generative adversarial network is as follows Figure 2 As shown;
[0071] Step S3 is specifically as follows:
[0072] S31: input the training 3D model into the generator to generate a predicted residual 3D model; downsample and upsample the label 3D model in sequence to generate a sampled 3D model, and subtract the label 3D model from the sampled 3D model to obtain a true residual 3D model;
[0073] S32: input the predicted residual three-dimensional model and the true residual three-dimensional model into the discriminator, and calculate and obtain the binary cross entropy loss function and the spatial context loss function;
[0074] Furthermore, the binary cross entropy loss function compares the probability distribution of the generated geological model with the distribution of the true label and calculates the distance between the generated model and the true label;
[0075] The calculation formula of the binary cross entropy loss function BCELoss is:
[0076]
[0077] Where i is the number of the training 3D model; N is the total number of training 3D models; y i is the true label of the true residual 3D model corresponding to the training 3D model i, and its value is 0 or 1; p i Represents the probability that the generated prediction residual 3D model is the true label.
[0078] Furthermore, the focus of the loss function is adjusted through the spatial context loss function, so that it pays more attention to the area related to the conditional profile data;
[0079] Spatial context loss function L ctext The calculation formula of (G) is:
[0080]
[0081] Among them, binary_mask(i) is the binary mask matrix, R real is the residual structure of the true residual 3D model, R generated is the residual structure of the predicted residual 3D model, Indicates that the training 3D model i obeys the real data distribution P data Expected sampling.
[0082] S33: Calculate the joint loss function through the binary cross entropy loss function and the spatial context loss function;
[0083] Furthermore, during the model generation process, a joint loss function is used to guide the network model to generate a more accurate geological model from both local and global aspects, ensuring that the training process can converge in the direction of respecting the conditional data while restoring the spatial heterogeneous patterns;
[0084] Joint loss function G * The calculation formula is:
[0085] G * =BCELoss+λ1L ctext (G)
[0086] Among them, BCELoss is the binary cross entropy loss function, L ctext (G) is the spatial context loss function, and λ1 is the loss coefficient.
[0087] S34: Back-propagate the generative adversarial network through the joint loss function to adjust the parameters of the generative adversarial network;
[0088] S35: Repeat steps S31-S34 until the joint loss function is less than a preset value to obtain a trained generative adversarial network.
[0089] S4: Input the three-dimensional model to be reconstructed into the trained generative adversarial network to obtain the reconstructed three-dimensional model.
[0090] Furthermore, in order to illustrate the feasibility of the method provided by the present invention, a three-dimensional experimental case of a binary discrete attribute and two multivariate discrete attributes was implemented according to the above steps.
[0091] like Figure 3 The figure shows a three-dimensional simulation experiment case of binary discrete attributes designed by this method. This experiment case uses a three-dimensional fold model with two fold structures as experimental data. Figure 3 (a) is a two-dimensional cross-sectional image extracted and cut from a large-scale three-dimensional model of 180×150×120 and used as conditional data. The conditional data needs to be filled into a three-dimensional model before use. Figure 3(b) is a small-scale 3D model cut from the original 3D model of 180×150×120 scale. It is a label set during the training process. These small-scale 3D models have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 3 (c) is a small-scale 3D model cut out from the padded 3D model. It is the training set during the training process. The dimensions of the training set also have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 3 (d) is the three-dimensional data model generated after the test data is reconstructed. This process is gradually generated from small scale to large scale, and the scales are 16×16×16, 32×32×32, and 64×64×64. From the reconstruction results of the binary discrete attribute three-dimensional model, it can be seen that the reconstructed test model has a similar spatial structure to the data model in the label set and has clear stripes, which meets the constraints of the conditional data.
[0092] Figure 4 The variogram and connectivity function curves of 20 experimental generation results and the reference model are shown. Figure 4 (b) shows the connectivity curve in both X and Y directions. Figure 4 (a) the variogram and Figure 4 It can be seen from the connectivity curve distribution of (b) that the generated curves of the two are very similar. Therefore, the statistical characteristics of the reconstructed model and the reference model show that the similarity between the two is very close, indicating that this method effectively restores the heterogeneous characteristics of the spatial structure in the three-dimensional reference model.
[0093] like Figure 5 The figure shows a three-dimensional simulation experiment case of multivariate discrete attributes designed by this method. This experiment case uses a stratum structure model with multivariate discrete attributes as the experimental data. The experimental data processing process is the same as that of the binary discrete attribute experiment. Figure 5 (a) is a two-dimensional cross-sectional image extracted and cut from a large-scale three-dimensional model of 228×114×114 and used as conditional data. The conditional data also needs to be filled into a three-dimensional model before use. Figure 5 (b) is a small-scale 3D model cut from the original 3D model of 228×114×114 scale. It is a label set during the training process. These small-scale 3D models have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 5 (c) is a small-scale 3D model cut out from the padded 3D model. It is the training set during the training process. The dimensions of the training set also have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 5(d) is the 3D data model generated after the test data is reconstructed. This process is gradually generated from small scale to large scale, and the scales are 16×16×16, 32×32×32, and 64×64×64. It can be seen from the multi-attribute 3D reconstruction results that the stratigraphic sedimentary facies with multiple different attributes in the reconstructed stratigraphic structure 3D model have obvious and clear stratification boundaries.
[0094] Figure 6 The variograms of 20 experimental generation results and the reference model were plotted to measure the similarity between the two. It can be seen from the variograms that the generation curves of the reference model and the simulated 3D model data have the same height and extension trend.
[0095] Figure 7 The figure shows a three-dimensional simulation experiment case of multivariate discrete attributes designed by this method. This experiment case uses a river phase reservoir model with multivariate discrete attributes as the experimental data. The experimental data processing process is the same as that of the binary discrete attribute experiment. Figure 7 (a) is a two-dimensional cross-sectional image extracted and cut from a large-scale three-dimensional model of 250×250×100 and used as conditional data. The conditional data also needs to be filled into a three-dimensional model before use. Figure 7 (b) is a small-scale 3D model cut from the original 3D model of 250×250×100. It is a label set during the training process. These small-scale 3D models have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 7 (c) is a small-scale 3D model cut out from the padded 3D model. It is the training set during the training process. The dimensions of the training set also have three different scales, namely 64×64×64, 32×32×32, and 16×16×16. Figure 7 (d) is the 3D data model generated after the test data is reconstructed. This process is gradually generated from small scale to large scale, and the scales are 16×16×16, 32×32×32, and 64×64×64. It can be seen from the multi-attribute 3D reconstruction results that the multiple channel facies with different attributes in the reconstructed 3D model of the channel facies reservoir structure have obvious and clear stratification boundaries.
[0096] Figure 8 The variograms of 20 experimental generation results and the reference model were plotted to measure the similarity between the two. It can be seen from the variograms that the generation curves of the reference model and the simulated 3D model data have the same height and extension trend.
[0097] The above three experimental cases show that the simulation results of the three-dimensional reconstruction method provided by the present invention and the reference model are visually similar to the statistical characteristics, proving that the present method successfully restores the spatial structural characteristics in the training data. This also confirms that the present method has the superior ability to accurately and quickly reconstruct complex three-dimensional geological structures under the constraints of multiple profile condition data.
[0098] See also Fig. 9 , Fig. 9 It is a working diagram of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a three-dimensional model reconstruction device 401 based on a generative adversarial network, a processor 402 and a storage medium 403.
[0099] A three-dimensional model reconstruction device 401 based on a generative adversarial network: The three-dimensional model reconstruction device 401 based on a generative adversarial network implements the three-dimensional model reconstruction method based on a generative adversarial network.
[0100] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the three-dimensional model reconstruction method based on the generative adversarial network.
[0101] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the three-dimensional model reconstruction method based on the generative adversarial network.
[0102] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0103] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0104] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A three-dimensional model reconstruction method based on a generative adversarial network, characterized in that: Includes steps: S1: obtaining an original three-dimensional model, and extracting multiple conditional sections from the original three-dimensional model; S2: Fill and segment each conditional section to obtain a set of training 3D models; segment the original 3D model to obtain a set of labeled 3D models; S3: construct a generative adversarial network, train the generative adversarial network by training the 3D model set and the label 3D model set, and obtain a trained generative adversarial network; S4: Input the three-dimensional model to be reconstructed into the trained generative adversarial network to obtain the reconstructed three-dimensional model.
2. The three-dimensional model reconstruction method based on generative adversarial network according to claim 1 is characterized in that: Step S2 is specifically as follows: S21: sorting the conditional sections according to their spatial positions in the original three-dimensional model, and filling the blank positions between the conditional sections to generate a filled three-dimensional model of the same size as the original three-dimensional model; S22: segmenting the filled three-dimensional model at different scales, and using the obtained filled three-dimensional sub-model sets at different scales as the training three-dimensional model set; S23: Segment the original three-dimensional model into different scales, and use the obtained original three-dimensional sub-model sets of different scales as the label three-dimensional model set.
3. The three-dimensional model reconstruction method based on generative adversarial network according to claim 1, characterized in that: Generative adversarial networks include: generator and discriminator; The generator includes a first SAGAN network, a first self-attention layer, a U-Net network, a second SAGAN network, and a second self-attention layer connected in sequence, and the first SAGAN network is connected to the second SAGAN network; The discriminator includes a first PatchGAN network, a second PatchGAN network, and a linear convolutional network connected in sequence.
4. The three-dimensional model reconstruction method based on generative adversarial network according to claim 3 is characterized in that: Step S3 is specifically as follows: S31: input the training 3D model into the generator to generate a predicted residual 3D model; downsample and upsample the label 3D model in sequence to generate a sampled 3D model, and subtract the label 3D model from the sampled 3D model to obtain a true residual 3D model; S32: input the predicted residual three-dimensional model and the true residual three-dimensional model into the discriminator, and calculate and obtain the binary cross entropy loss function and the spatial context loss function; S33: Calculate the joint loss function through the binary cross entropy loss function and the spatial context loss function; S34: Back-propagate the generative adversarial network through the joint loss function to adjust the parameters of the generative adversarial network; S35: Repeat steps S31-S34 until the joint loss function is less than a preset value to obtain a trained generative adversarial network.
5. The three-dimensional model reconstruction method based on generative adversarial network according to claim 4, characterized in that: The calculation formula of the binary cross entropy loss function BCELoss is: Where i is the number of the training 3D model; N is the total number of training 3D models; y i is the true label of the true residual 3D model corresponding to the training 3D model i, and its value is 0 or 1; p i Represents the probability that the generated prediction residual 3D model is the true label.
6. The three-dimensional model reconstruction method based on generative adversarial network according to claim 4, characterized in that: Spatial context loss function L ctext The calculation formula of (G) is: Among them, binary_mask(i) is the binary mask matrix, R real is the residual structure of the true residual 3D model, R generated is the residual structure of the predicted residual 3D model, Indicates that the training 3D model i obeys the real data distribution P data Expected sampling.
7. The three-dimensional model reconstruction method based on generative adversarial network according to claim 4, characterized in that: Joint loss function G * The calculation formula is: G * =BCELoss+λ1L ctext (G) Among them, BCELoss is the binary cross entropy loss function, L ctext (G) is the spatial context loss function, and λ1 is the loss coefficient.
8. A storage medium, characterized in that: The storage medium stores instructions and data for implementing the three-dimensional model reconstruction method based on a generative adversarial network as described in any one of claims 1 to 7.
9. A three-dimensional model reconstruction device based on a generative adversarial network, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the three-dimensional model reconstruction method based on a generative adversarial network as described in any one of claims 1 to 7.