A method for training a super-resolution generator of time-varying data

Abstract

The application relates to a training method of a time-varying volume data super-resolution generator, comprising the following steps: S1, obtaining original volume data; S2, performing clipping to obtain real high-resolution volume data; S3, performing three-linear down-sampling to obtain real low-resolution volume data; S4, performing splicing in the channel dimension to obtain first splicing volume data and inputting the first splicing volume data into a generator to generate fake high-resolution volume data; S5, splicing the real high-resolution volume data, the fake high-resolution volume data and the size-adjusted real low-resolution volume data in the channel dimension to obtain second splicing volume data and third splicing volume data respectively, and inputting the second splicing volume data and the third splicing volume data into a discriminator to obtain a first prediction matrix and a second prediction matrix respectively; S6, training the generator and the discriminator based on the volume data and the prediction matrix; and S7, repeating steps S1-S6 until a preset iteration number is reached. The application can obtain a generator with good time-varying volume data super-resolution generation effect through adversarial training.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of super-resolution reconstruction, and particularly relates to a training method of a super-resolution generator of time-varying volume data. BACKGROUND

[0002] Time-varying volume data often produces in scientific simulation to be applied to various researches, such as weather forecast, computational fluid dynamics, combustion science, computational cosmology, climate model research, etc. These time-varying volume data usually contains complex and large-scale features, and needs to be applied to in-situ visualization technology.

[0003] In-situ visualization means that the feature data is first reduced during simulation to reduce the pressure of reading, writing and storage, and then the data is up-sampled in the post-processing stage of analysis and visualization. The up-sampling technology, i.e. super-resolution, plays an important role in in-situ visualization.

[0004] Although a large number of up-sampling technologies have been proposed, these researches rarely focus on time-varying volume data. In other words, given a sequence of low-resolution time-varying volume data, we want to map it into a sequence of high-resolution time-varying volume data, which brings two challenges. Compared with the past technology, the time coherence of time-varying volume data must be considered, and compared with the traditional method, our method wants to achieve better visual effect. SUMMARY

[0005] In view of the above problems existing in the prior art, the application provides a training method of a super-resolution generator of time-varying volume data, which trains the super-resolution generator by setting a discriminator and by mutual confrontation of the discriminator and the generator, so as to obtain a generator with good time-varying volume data super-resolution generation effect, and lays a foundation for weather forecast, computational fluid dynamics, combustion science, computational cosmology and climate model research.

[0006] The application adopts the following technical scheme:

[0007] The training method of the super-resolution generator of the time-varying volume data comprises the following steps:

[0008] S1, selecting time-varying volume data continuous at multiple time steps in a training set to obtain original volume data;

[0009] S2, cropping the original volume data to obtain real high-resolution volume data;

[0010] S3, performing three-linear down-sampling on the real high-resolution volume data to obtain real low-resolution volume data;

[0011] S4, splicing the real low-resolution volume data in the channel dimension to obtain first spliced volume data and inputting the generator to generate the fake high-resolution volume data;

[0012] S5, splicing the real high-resolution volume data, the fake high-resolution volume data, and the size-adjusted real low-resolution volume data in the channel dimension to obtain second spliced volume data and third spliced volume data, respectively, and inputting the discriminator to obtain a first prediction matrix and a second prediction matrix, respectively;

[0013] S6, calculating a generator training loss function value based on the real high-resolution volume data, the fake high-resolution volume data, the second prediction matrix, and the intermediate features of the second spliced volume data in the discriminator, and the intermediate features of the third spliced volume data in the discriminator, and calculating a discriminator training loss function value based on the first prediction matrix and the second prediction matrix;

[0014] S7, training the generator based on the generator training loss function value to obtain a trained generator, and training the discriminator based on the discriminator training loss function value to obtain a trained discriminator;

[0015] S8, repeating steps S1-S7 until a preset number of iterations is reached.

[0016] As a preferred scheme, step S1 is specifically: selecting 3 time-varying volume data continuous in time steps in the training set to obtain the original volume data.

[0017] As a preferred scheme, step S2 is specifically: cropping each data in the original volume data at the same position to obtain the real high-resolution volume data.

[0018] As a preferred scheme, the generator in step S4 includes a first encoder and a decoder, the first encoder includes a first encoding block and a second encoding block connected in sequence, the decoder includes a first decoding block, a second decoding block, a third decoding block, and a fourth decoding block connected in sequence, and the second encoding block is connected with the first decoding block.

[0019] A plurality of skip links are formed between the encoder and the decoder to flow the early features lost in the first encoder into the decoding block.

[0020] As a preferred scheme, the discriminator in step S5 includes a second encoder, and the second encoder includes a third encoding block, a fourth encoding block, and a fifth encoding block connected in sequence.

[0021] As a preferred scheme, the calculation formula of the generator training loss function value in step S6 is:

[0022] L(θ G )=λ adv ×Ladv +λ voxdis ×L voxdis +λ featdis ×L featdis ,

[0023] wherein, L adv represents an adversarial loss, L adv is calculated based on the second prediction matrix; L voxdis represents a voxel distance loss, L voxdis is calculated based on the real high-resolution volume data and the fake high-resolution volume data; L featdis represents a feature loss, L featdis is calculated based on the intermediate features of the second spliced volume data and the third spliced volume data in the discriminator; λ adv , λ voxdis , λ featdis are weights of L adv , L voxdis , L voxdis respectively.

[0024] As a preferred solution, the calculation formula of the adversarial loss L adv is as follows:

[0025]

[0026] wherein, BCE represents a binary cross-entropy loss, D represents a discriminator, G represents a generator, represents real low-resolution volume data, represents channel concatenation, R represents a correction size operation, represents a second prediction matrix, and 1 represents a matrix with values all being 1.

[0027] As a preferred solution, the calculation formula of the voxel distance loss L voxdis is as follows:

[0028]

[0029] wherein, ||·|| represents an L1 loss, L1 represents real high-resolution volume data, represents fake high-resolution volume data.

[0030] As a preferred solution, the calculation formula of the feature loss L featdis is as follows:

[0031]

[0032] wherein, MSE represents a mean squared error loss, ​respectively represent the intermediate features of the second spliced data and the third spliced data in the lth layer of the discriminator.

[0033] As a preferred solution, the calculation formula of the discriminator training loss function value in step S6 is:

[0034]

[0035] wherein, BCE represents the binary cross-entropy loss, D represents the discriminator, G represents the generator, represents the real low-resolution volume data, represents the channel splicing, and R represents the correction size operation, represents the second prediction matrix, and 0 represents a matrix with all values being 0, represents the real high-resolution volume data, represents the first prediction matrix.

[0036] The beneficial effects of the present application are:

[0037] By setting a discriminator, and training the super-resolution generator through the mutual confrontation between the discriminator and the generator, a generator with good super-resolution generation effect of time-varying volume data is obtained, which lays a foundation for weather forecasting, computational fluid dynamics, combustion science, computational cosmology, and climate model research. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0039] Figure 1 is a flow chart of the training method of the super-resolution generator of time-varying volume data according to the present application;

[0040] Figure 2 is a structural schematic diagram of the generator according to the present application;

[0041] Figure 3 is a structural schematic diagram of the decoder according to the present application;

[0042] Figure 4 is a comparison diagram of the super-resolution reconstruction of the SquareCylinder data set by the generator trained by the training method according to the present application and the original data;

[0043] Figure 5is a comparison chart of the Ionization (H) data set after super-resolution reconstruction by the generator trained by the training method of the present application and the original data. DETAILED DESCRIPTION

[0044] The present application will be described in more detail by the following specific embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the specification. The present application can also be implemented or applied by other different embodiments, and various modifications or changes can be made to the details in the specification based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0045] Let be a sequence of low-resolution volume data in the time-varying volume data set, where, are low-resolution volume data at time steps 1, 2,..., n, respectively; let be a sequence of high-resolution volume data in the time-varying volume data set, where, are high-resolution volume data at time steps 1, 2,..., n, respectively. The present application aims to obtain the mapping relationship from V l to V h , i.e. V h = M(V l ), while considering spatial coherence and temporal coherence.

[0046] Referring to FIG. Figure 1 , the present embodiment provides a training method of a super-resolution generator of time-varying volume data, comprising the steps of:

[0047] S1, selecting a plurality of time steps of time-varying volume data in the training set to obtain original volume data V

[0048] S2, randomly selecting a position, and performing the same cropping on the original volume data V o at the position to obtain real high-resolution volume data V

[0049] S3, performing trilinear downsampling on the real high-resolution volume data V to obtain real low-resolution volume data V

[0050] S4, concatenating the real low-resolution volume data V in the channel dimension to obtain first concatenated volume data and inputting the first concatenated volume data into the generator G to generate fake high-resolution volume data V Still a whole, the internal data is spliced together in the channel dimension;

[0051] S5, real high-resolution volume data fake high-resolution volume data respectively with the size-adjusted real low-resolution volume data spliced in the channel dimension to obtain second spliced volume data and third spliced volume data respectively, and input into the discriminator D to obtain first prediction matrix and second prediction matrix respectively;

[0052] S6, based on the real high-resolution volume data fake high-resolution volume data The second prediction matrix and the intermediate features of the second spliced volume data in the discriminator and the intermediate features of the third spliced volume data in the discriminator are used to calculate the generator training loss function value, and the discriminator training loss function value is calculated based on the first prediction matrix and the second prediction matrix; the second prediction matrix is passed to the generator to guide the generator to generate as close as possible This constitutes a competitive cycle between the generator and the discriminator.

[0053] S7, based on the generator training loss function value, the generator is trained to obtain a trained generator, and based on the discriminator training loss function value, the discriminator is trained to obtain a trained discriminator;

[0054] S8, repeat steps S1-S7 until a preset iteration number is reached;

[0055] S9, select a plurality of time-varying volume data in the test set in a continuous time step, input into the trained generator after three linear downsampling and channel splicing operation to generate corresponding fake high-resolution volume data, and perform visualization operation after disassembling, judge the training effect of the generator through the visualization animation, if the expected is reached, stop training, if the expected is not reached, return to step S1 and continue training.

[0056] It should be noted that the time-varying volume data represents a series of three-dimensional data continuous in time, which can be hurricane simulation time-varying volume data, salt dissolution behavior simulation time-varying volume data, hydrogen ion ionization instability simulation time-varying volume data, and incompressible flow field simulation time-varying volume data, etc.

[0057] Referring to Figure 1 , the generative adversarial network in our training method includes a generator and a discriminator. The task of the generator is to generate fake high-resolution volume data as close as possible to the real high-resolution volume data and deceive the discriminator. The task of the discriminator is to distinguish ​With and feed back to the generator. Unlike the traditional generative adversarial network, which is only composed of a simple convolutional neural network, the generator and discriminator in our model are composed of more complex neural networks inspired by the Pix2Pix model. Therefore, we call our model Pix2PixSSR.

[0058] In particular, the following describes the architecture of the generator and discriminator in the present application and the loss function calculation formula for training them:

[0059] Referring to Figure 2 , the architecture of the generator in the present application is shown. k represents the core size, #k represents the number of cores, p represents the padding value, s represents the step, sf represents the sampling rate, BN represents batch normalization, DO represents dropout, LReLU represents leaky rectified linear unit, conv represents convolution, up represents upsampling, concat represents concatenation, and maxpool represents maximum pooling.

[0060] Our generator contains an encoder-decoder architecture. The task of the encoder is to encode and extract the features of the input real low-resolution volume data, and store these features in the intermediate data containing high-level and important features of small size. The encoder contains two encoding blocks, named encoding block 1 and encoding block 2 respectively. Each encoding block contains two convolutional layers and one max-pooling layer. The task of the decoder is to amplify the features of the intermediate data, so as to obtain the features of the fake high-resolution volume data. The decoder contains four decoding blocks, named decoding block 1, decoding block 2, decoding block 3 and decoding block 4 respectively. Each decoding block is composed of an upsampling-convolution- concatenation structure, and is attached with two convolutional layers.

[0061] In addition to the encoder-decoder architecture, our generator also has four skip links, which allow features to flow from the encoding blocks to the decoding blocks. The benefits of skip links are twofold. For forward propagation, it allows early features lost in the encoding blocks to flow into the decoding blocks, resulting in better detail effects. For backpropagation, it can avoid the problem of gradient vanishing.

[0062] Our generator consists of seven operation steps, which are convolution layer 0 operation, up-sampling operation, convolution layer 1 operation, max-pooling operation, up-sampling convolution concatenation operation, convolution 2 operation, and leaky rectified linear unit operation. All the operation steps are running on three-dimensional data, such as 24x8x6x3 (widthxheightxdepthxchannel), thus are three-dimensional operations. In brief, we use, for example, k=2, s=2, p=2, sf=2 to represent the size of kernel, the size of stride, the size of padding, and the sampling ratio are the same in all dimensions (x, y, z). When their sizes are different in different dimensions, we use, for example, k=(2, 2, 1) to represent.

[0063] Convolution 0 operation is a convolution operation with kernel size (1, 1, 1). Convolution 1 operation is a convolution operation with kernel size (3, 3, 3) with batch normalization, dropout, and leaky rectified linear unit. Convolution 2 is a convolution operation with kernel size (3, 3, 3) with batch normalization and rectified linear unit. Up-sampling operation increases the three-dimensional data in three dimensions to twice or four times of the original. Max-pooling operation reduces the three-dimensional data in three dimensions to one half of the original. Up-sampling convolution concatenation operation implements the skip connection, which is a hybrid sequence of up-sampling, convolution, and concatenation operation. For example, the underlying data 6x2x3x512 in Figure 2

[0064] ​A low-resolution 3D data input generator, first through the encoding block 1, through three convolution 0 operations, the number of channels is enlarged from 3 to 128, and then through two convolution 1 operations, the number of channels is enlarged to 256, through a maximum pooling operation, the data is reduced by half in three dimensions, and enters the encoding block 2. The encoding block 2 includes two convolution 1 operations, and the number of channels is enlarged to 512. After a maximum pooling operation, the data is reduced by half in three dimensions. After another upsampling convolution splicing operation, the data is enlarged by one time in three dimensions, and enters the decoding block 1. The decoding block 1 includes two convolution 2 operations and an upsampling convolution splicing operation, and the number of channels is reduced from 512 to 256, and enlarged by one time in three dimensions. Enter the decoding block 2, which includes two convolution 2 operations and an upsampling convolution splicing operation, and the number of channels is reduced to 128, and the data is enlarged by one time in three dimensions. Enter the decoding block 3, which includes two convolution 2 operations and an upsampling convolution splicing operation, and the number of channels is reduced to 64, and the data is enlarged by one time in three dimensions. At this time, the size of the data in three dimensions is each 4 times the original input size, and enters the decoding block 4. The decoding block 4 includes two convolution 2 operations, and the number of channels is reduced to 32, and then through a convolution 0, the number of channels is reduced to 3, and finally through a leaky rectified linear unit operation, the final output is obtained. The number of channels of the final output is consistent with the input data, and the volume is 64 times that of the original data.

[0065] In the generator, four skip links are also included, which connect the data of the early encoding block to the later decoding block, allowing features to flow from the encoding block to the decoding block. For forward propagation, it enables early features lost in the encoding block to flow into the decoding block, thereby achieving better detail effects. For backward propagation, it can avoid the problem of gradient disappearance. The first skip link directly quadruples the second data in the encoding block 1 and links it to the decoding block 4. The second skip link doubles the third data in the encoding block 1 and links it to the decoding block 3. The third skip link links the fourth data in the encoding block 1 to the decoding block 2. The fourth skip link links the first data in the encoding block 2 to the decoding block 1.

[0066] Referring to Figure 3 As shown in the figure, the discriminator architecture in the present application is shown, which is a typical PatchGAN model and only has an encoder. The encoder includes three encoding blocks, which are named encoding block 1, encoding block 2 and encoding block 3. The encoding block is consistent with that in the generator, except that there is no discard operation in the convolution 1 operation. A common problem of GAN is that the discriminator often has more advantages than the generator, because the task of the discriminator is much easier than that of the generator. In order to avoid this problem, we maintain the simplicity of the discriminator architecture, compared with our generator architecture.

[0067] In the discriminator, a three-dimensional input data first goes through a convolution 0 operation, changing the number of channels from 6 to 32, then goes through two convolution 1 operations, changing the number of channels to 64, and then goes through a max-pooling operation, reducing the data by half in three dimensions. This is the encoding block 1. The data enters the encoding block 2, where it goes through two convolution 1 operations, changing the number of channels to 128, and then goes through a max-pooling operation, reducing the data by half in three dimensions. The data enters the encoding block 3, where it goes through two convolution 1 operations, changing the number of channels to 256, and then goes through a max-pooling operation, reducing the data by half in three dimensions. Finally, a convolution 0 operation is performed, changing the number of channels to 1, to obtain the final output. The output of the discriminator will be used as a feedback data input to the generator to guide the training of the generator.

[0068] Referring to Figure 2 As shown, we use dropout operation in the encoding block 1 and the encoding block 2 in the discriminator. Dropout is a regularization technique in which some neurons in each layer are randomly turned off at each iteration to temporarily eliminate their influence on downstream neuron activations in forward propagation, and any weight updates are not applied to neurons in backpropagation. Dropout makes our generator less sensitive to specific weights of neurons, so it can better generalize and avoid overfitting. Referring to Figure 2 , Figure 3 As shown in

[0069] Generator loss function:

[0070] Let G and D be the generator and the discriminator, respectively, and θ G and θ D be the learning parameters of G and D. The formula for calculating the generator training loss function value is:

[0071] L(θ G )=λ adv ×L adv +λ voxdis ×L voxdis +λ featdis ×L featdis ,

[0072] where L adv represents the adversarial loss, L adv is calculated based on the second prediction matrix; L voxdix represents the voxel distance loss, and L voxdisCalculated based on real high-resolution volume data and simulated high-resolution volume data; L featdis L represents the feature loss. featdis λ is calculated based on the intermediate features of the second and third splice data in the discriminator. adv , λ voxdis , λ featdis L respectively adv , λ voxdis , λ voxdis The weight.

[0073] The three losses mentioned above are detailed below:

[0074] The first loss is the adversarial loss L. adv The calculation formula is:

[0075]

[0076] Where BCE represents the binary cross-entropy loss, D represents the discriminator, and G represents the generator. Represents true low-resolution volumetric data. This indicates channel splicing; R indicates size correction operation. 1 represents the second prediction matrix, and 1 represents a matrix where all values ​​are 1.

[0077] High-resolution volume data faked by the generator Compared to real low-resolution volume data with the same dimensions as the former, the data was corrected. Simultaneously, the data is fed into a discriminator. The discriminator doesn't know beforehand which data is real and which is fake. It calculates and outputs a corresponding probability value for each input to determine whether the data is real or fake. The binary classification cross-entropy loss is calculated from the output probability values, yielding the adversarial loss L. adv , antagonistic loss L adv This will be fed back to the generator.

[0078] The second loss is called the voxel distance loss L. voxdis The calculation formula is:

[0079]

[0080] Among them, ||·|| L1 This represents L1 loss, which serves two purposes: first, it allows the generator to see the true values, thus enabling it to generate higher-resolution volumetric data that more closely approximates reality; second, using L1 loss can improve training stability. Represents true high-resolution volume data. This indicates falsified high-resolution volumetric data, with a voxel distance loss L. voxdisThe real high-resolution volume data is compared with the fake high-resolution volume data , and the L1 loss is calculated, which reflects the difference between the two. voxdis The feedback will be given to the generator.

[0081] The third loss is the feature loss L featdis , which describes the feature distance between the second spliced data V′ r and the third spliced data V′ f , and the calculation formula is:

[0082]

[0083] where MSE represents the average variance loss, and V′ and V′

[0084] represent the lth intermediate feature of the second spliced data and the third spliced data in the discriminator, respectively. represent the lth intermediate feature of the second spliced data and the third spliced data in the discriminator, respectively. r f MSE represents the average variance loss. This loss aims to make the generator produce high-resolution volume data closer to the real value. The feature loss L featdis will be fed back to the generator.

[0085] Discriminator loss function:

[0086] Equation 5 shows the loss function L(θ D ) of the discriminator, which contains two adversarial losses. Similar to the adversarial loss in the generator, the first loss represents the prediction of fake data, and the second loss represents the prediction of real data. The average of the two losses is the total loss of the discriminator. Since the output of the discriminator is a matrix, 0 and 1 here represent the matrix with all 0 values and the matrix with all 1 values, respectively. The loss function of the discriminator will be fed back to the discriminator to help the discriminator train. The calculation formula of the discriminator training loss function value is:

[0087]

[0088] where BCE represents the binary cross-entropy loss, D represents the discriminator, G represents the generator, represents the real low-resolution volume data, represents channel concatenation, R represents a resize operation, represents a second prediction matrix, 0 represents a matrix with all values being 0, represents real high-resolution volume data, represents a first prediction matrix.

[0089] The visualization operation described in the present application is explained as follows:

[0090] We use the volume ray casting algorithm in volume rendering to present the time-varying high-resolution volume data inferred by the trained generator in the form of animation step by step, so that we can observe their evolution. In particular, in order to emphasize the boundary of the foreground feature of the volume data, we use the following formula to set the opacity of their transfer function:

[0091] O(v f )=normInten(v f )×β+log(normGrad(v f )+1)×γ,

[0092] where v f represents the voxel of each foreground feature, O(v f ) represents the opacity; normInten(v f ) and normGrad(v f ) represent the normalized intensity and gradient, respectively; log(·) represents the logarithmic function; β and γ represent two user-controlled parameters that control the intensity and boundary opacity, respectively. In order to visualize the inferred volume data, we use volume rendering to visualize the difference between the inferred volume data and the real volume data, and present it in the form of animation, so as to better identify the difference, so as to judge the training effect of the generator, so as to ensure the effect of the generator on the super-resolution generation of time-varying volume data.

[0093] The effect of the generator on the super-resolution generation of time-varying volume data based on the internal structure of the generator and the training method described in the present application after training can be referred to as shown in Figure 4 、 Figure 5 , Figure 4 is a comparison chart of the generator trained by the training method described in the present application for super-resolution reconstruction of the SquareCylinder data set and the original data, Figure 5 is a comparison chart of the generator trained by the training method described in the present application for super-resolution reconstruction of the Ionization(H) data set and the original data.

[0094] The above-described embodiments are merely intended to describe the preferred embodiments of the present application, and are not intended to limit the scope of the present application. Various modifications and improvements to the present application made by those skilled in the art are intended to fall within the scope of the present application.

Claims

1. A method of training a super-resolution generator of time-varying data, characterized in that, The method comprises the steps of: S1, selecting a plurality of time step continuous time-varying body data in a training set to obtain original body data; S2, cropping the original body data to obtain real high-resolution body data; S3, performing three-linear downsampling on the real high-resolution body data to obtain real low-resolution body data; S4, splicing the real low-resolution body data in the channel dimension to obtain first spliced body data and inputting the generator to generate fake high-resolution body data; S5, splicing the real high-resolution body data, the fake high-resolution body data and the real low-resolution body data of the size adjusted real low-resolution body data in the channel dimension to obtain second spliced body data and third spliced body data, respectively, and inputting the discriminator to obtain first prediction matrix and second prediction matrix, respectively; S6, calculating a generator training loss function value based on the real high-resolution body data, the fake high-resolution body data, the second prediction matrix, the intermediate features of the second spliced body data in the discriminator and the intermediate features of the third spliced body data in the discriminator, and calculating a discriminator training loss function value based on the first prediction matrix and the second prediction matrix; S7, training the generator based on the generator training loss function value to obtain a trained generator, and training the discriminator based on the discriminator training loss function value to obtain a trained discriminator; S8, repeating steps S1-S7 until a preset iteration number is reached; The generator in step S4 comprises a first encoder and a decoder, the first encoder comprises a first encoding block and a second encoding block connected in sequence, the decoder comprises a first decoding block, a second decoding block, a third decoding block and a fourth decoding block connected in sequence, and the second encoding block is connected with the first decoding block; A plurality of skip links are formed between the encoder and the decoder to make early features lost in the first encoder flow into the decoding block; The calculation formula of the generator training loss function value in step S6 is: ， wherein, represents an adversarial loss, is calculated based on the second prediction matrix; represents a voxel distance loss, is calculated based on the real high-resolution volume data and the fake high-resolution volume data; represents a feature loss, is calculated based on the intermediate features of the second spliced volume data and the third spliced volume data in the discriminator; , , are respectively , , weights of After step S8, the method further comprises: using a volume ray projection algorithm in volume rendering to present the time-varying high-resolution body data inferred by the trained generator in the form of animation.

2. The method of claim 1, wherein, Step S1 specifically comprises: selecting three time step continuous time-varying body data in a training set to obtain original body data.

3. The method of claim 1, wherein, Step S2 specifically comprises: cropping each data in the original body data at the same position to obtain real high-resolution body data.

4. The method of claim 1, wherein, The discriminator in step S5 comprises a second encoder, and the second encoder comprises a third encoding block, a fourth encoding block and a fifth encoding block connected in sequence.

5. The method of claim 1, wherein, Adversarial loss The calculation formula is: ， Where BCE represents binary cross-entropy loss, D represents the discriminator, G represents the generator, represents real low-resolution volume data, represents channel concatenation, R represents a resize operation, represents a second prediction matrix, and 1 represents a matrix with all values being 1.

6. The method of claim 1, wherein, voxel distance loss The formula for calculating is: ， wherein, denotes the LI loss, denotes the real high-resolution volume data, denotes the fake high-resolution volume data.

7. The method of claim 1, wherein, Characteristic loss The calculation formula is: ， Where MSE represents the average variance loss, , These represent the second and third concatenated data in the discriminator, respectively. Intermediate features of the layer.

8. The method of claim 1, wherein, The calculation formula of the discriminator training loss function value in step S6 is: ， where BCE denotes binary cross-entropy loss, D denotes the discriminator, and G denotes the generator, denotes real low-resolution volume data, denotes channel concatenation, and R denotes a resize operation, denotes a second prediction matrix, and 0 denotes a matrix of all zeros, denotes real high-resolution volume data, denotes a first prediction matrix.

Images