Intelligent virtual fitting method based on three-dimensional human body modeling

Abstract

The invention discloses an intelligent virtual fitting method based on three-dimensional human body modeling, which is applied to the technical field of virtual reality, and comprises the following steps: obtaining two-dimensional images of a user body and clothes to be fitted, adjusting and synthesizing a three-dimensional image based on perspective transformation, and extracting three-dimensional features by using a three-dimensional convolutional neural network; based on the three-dimensional features, performing three-dimensional text description prediction by using a bidirectional long-short-term memory network, and outputting a three-dimensional fitting image by using a three-dimensional deconvolution neural network in combination with the three-dimensional features; judging the authenticity of a three-dimensional fitting image based on the combination of three-dimensional convolution and a full-connection layer, calculating a cross entropy to determine a first loss, performing three-dimensional text description prediction after three-dimensional fitting by using a bidirectional long-short-term memory network, calculating a second loss based on term frequency-inverse document frequency, and performing three-dimensional text description prediction after three-dimensional fitting. And three-dimensional features after three-dimensional fitting are extracted based on the three-dimensional convolutional neural network, third loss is calculated through the Euclidean distance, and iterative parameter adjustment is performed respectively. According to the invention, the three-dimensional virtual fitting effect and precision are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of virtual reality technology, and in particular to an intelligent virtual fitting method based on three-dimensional human body modeling. Background Art

[0002] With the popularity of e-commerce platforms, the demand for online clothing shopping has increased significantly, and the network virtual fitting technology based on 3D image processing has also shown great market potential. Through virtual fitting, users can simulate the effect of clothes on virtual characters of the same body shape as the person fitting, thereby helping consumers choose the most suitable clothes. This not only effectively improves the user satisfaction of online clothing shopping, but also reduces the return rate of online clothing shopping, thereby further reducing the cost of online clothing shopping and promoting the development of online clothing shopping.

[0003] At present, the most common virtual fitting technology is based on the three-dimensional image of the user's body and the clothes to be tried on, and uses the three-dimensional feature extraction based on the neural network and the virtual fitting three-dimensional image generation model to perform the corresponding three-dimensional fitting imaging. Although it produces a certain three-dimensional visual effect, the high-precision requirements for the virtual fitting effect still need to be improved. In addition, in the above virtual fitting technology, the process of obtaining the three-dimensional image is relatively cumbersome and the accuracy is insufficient, which further affects the subsequent virtual fitting effect.

[0004] Therefore, how to provide an intelligent virtual fitting method based on three-dimensional human body modeling that can effectively improve the virtual fitting effect is a problem that those skilled in the art urgently need to solve. Summary of the invention

[0005] In view of this, the present invention proposes an intelligent virtual fitting method based on three-dimensional human body modeling.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] An intelligent virtual fitting method based on three-dimensional human body modeling, comprising:

[0008] Step 1: Obtain two-dimensional images of the user's body and the clothes to be tried, and adjust and synthesize three-dimensional images of the user's body and the clothes to be tried based on perspective transformation, and use a three-dimensional convolutional neural network to extract three-dimensional features of the user's body and the clothes to be tried;

[0009] Step 2: Based on the 3D features, a bidirectional long short-term memory network is used to predict the 3D text description of the user's body and the clothes to be tried on, and combined with the 3D features, a 3D deconvolutional neural network is used to output a 3D fitting image;

[0010] Step 3: Determine the authenticity of the three-dimensional fitting image based on the combination of three-dimensional convolution and fully connected layers, determine the first loss by calculating the cross entropy, and iteratively adjust the parameters of the three-dimensional convolutional neural network; locate the user body and the clothes to be tried on in the three-dimensional fitting image based on the target detection model, and use the bidirectional long short-term memory network to predict the three-dimensional text description of the user body and the clothes to be tried on after the three-dimensional fitting, calculate the second loss based on the term frequency-inverse document frequency, and iteratively adjust the parameters of the bidirectional long short-term memory network; extract the three-dimensional features of the user body and the clothes to be tried on after the three-dimensional fitting based on the three-dimensional convolutional neural network, calculate the third loss by the Euclidean distance, and iteratively adjust the parameters of the three-dimensional convolutional neural network; when the weighted sum of the first loss, the second loss, and the third loss is less than the preset loss threshold, the final intelligent virtual fitting model is obtained.

[0011] Optionally, in step 1, the three-dimensional image of the user's body and the clothing to be tried is adjusted and synthesized based on the perspective transformation, specifically:

[0012] The two-dimensional images of the user's body and the clothes to be tried are perspective transformed to obtain left and right eye images, and the distance between the left and right eye images is adjusted, and the left and right eye images after the distance adjustment are synthesized to obtain a three-dimensional image of the user's body and the clothes to be tried.

[0013] Optionally, perspective transformation is performed on the two-dimensional images of the user's body and the clothing to be tried to obtain left and right eye images, specifically:

[0014] Extracting the size of the two-dimensional image, scanning the two-dimensional image line by line, and linearly scaling the image of each line in turn to obtain a first image;

[0015] Perform mirror processing on the first image to obtain a second image, and use the first image and the second image as left-eye and right-eye images.

[0016] Optionally, adjust the distance between the left and right eye images, specifically:

[0017] One of the left and right eye images is displaced horizontally so that the distance between the left and right eye images reaches a preset distance to create binocular parallax and convergence angle.

[0018] Optionally, in step 1, the three-dimensional convolutional neural network is composed of a plurality of three-dimensional feature processing units connected in sequence, each of which includes: a three-dimensional convolution block, an activation layer, and a three-dimensional pooling layer, and each three-dimensional convolution block includes a plurality of three-dimensional convolution kernels; the three-dimensional convolution block is used to perform three-dimensional convolution processing on the input image to extract three-dimensional convolution features; the activation layer is used to add nonlinear features to the three-dimensional convolution features; and the pooling layer is used to reduce the dimensional size of the nonlinear features;

[0019] The output of the nth 3D convolution kernel in the i-th 3D feature processing unit among the multiple 3D feature processing units is as follows:

[0020]

[0021] Among them, F1 is the activation function used by the i-th three-dimensional feature processing unit; T (i-1)m (x+u, y+v, z+s) is the output of the mth 3D convolution kernel in the i-1th 3D feature processing unit;

[0022] is the sparse matrix corresponding to the nth three-dimensional convolution kernel in the i-th three-dimensional feature processing unit;

[0023] b in is the bias term of the i-th 3D feature processing unit; u, v, s are the sampling steps in the x, y, and z directions.

[0024] Optionally, in step 2, the prediction based on the three-dimensional text description of the user's body and the clothing to be tried is combined with the three-dimensional features, specifically:

[0025] The three-dimensional features of the user's body and the corresponding feature matrix and feature vector in the three-dimensional text description, as well as the three-dimensional features of the clothes to be tried and the corresponding feature matrix and feature vector in the three-dimensional text description are multiplied respectively, and the feature matrices and feature vectors corresponding to the two features obtained after the multiplication are concatenated as the input of the three-dimensional deconvolutional neural network.

[0026] Optionally, in step 2, the three-dimensional deconvolution neural network is composed of a plurality of three-dimensional restoration processing units connected in sequence, and each three-dimensional restoration processing unit includes: a three-dimensional transposed convolution block, an activation layer, and an anti-pooling layer; the three-dimensional transposed convolution block is used to perform three-dimensional deconvolution processing on the input image to obtain a three-dimensional deconvolution restoration image; the activation layer is used to add nonlinear features to the three-dimensional transposed convolution; and the anti-pooling layer is used to restore features.

[0027] Optionally, in step 3, the authenticity of the three-dimensional fitting image is determined based on the combination of the three-dimensional convolution and the fully connected layer, specifically:

[0028] It consists of multiple three-dimensional feature restoration units connected in sequence and a fully connected layer; each three-dimensional feature restoration unit includes: a three-dimensional convolution block, an activation layer, and a discard layer; the fully connected layer is used to make authenticity judgments based on the final outputs of multiple three-dimensional feature restoration units connected in sequence.

[0029] Optionally, in step 3, based on the authenticity of the three-dimensional fitting image, the cross entropy is calculated to determine the first loss as follows:

[0030] L 1 =Ex [logD(x)]+E x,t [log(1-D(G(x,t)))];

[0031] Wherein, D(x) is the result of determining the authenticity of the three-dimensional fitting image; and D(G(x, t)) is the result of determining the authenticity of the three-dimensional image of the user's body.

[0032] Optionally, in step 3, the third loss is calculated by Euclidean distance as follows:

[0033]

[0034] Among them, H is the number of three-dimensional features; K is the dimension of the three-dimensional features; x 1k The three-dimensional features before three-dimensional fitting and the three-dimensional features after three-dimensional fitting.

[0035] It can be seen from the above technical solutions that compared with the prior art, the present invention proposes an intelligent virtual fitting method based on three-dimensional human body modeling. Based on the extraction of three-dimensional features of three-dimensional images at different scales by three-dimensional convolutional neural networks, the present invention introduces three-dimensional text descriptions using bidirectional long short-term memory network predictions, and multiplies and splices fusions to jointly perform three-dimensional fitting prediction outputs, effectively improving the three-dimensional virtual fitting effect and accuracy. In addition, the authenticity of the three-dimensional fitting image is determined by combining the three-dimensional convolution and the fully connected layer, and the first loss is determined by calculating the cross entropy, and the three-dimensional text description of the user's body and the clothes to be tried after three-dimensional fitting is predicted based on the bidirectional long short-term memory network, and the second loss is calculated based on the term frequency-inverse document frequency, and the three-dimensional features of the user's body and the clothes to be tried after three-dimensional fitting are extracted based on the three-dimensional convolutional neural network, and the third loss is calculated by the Euclidean distance, and the three-dimensional convolutional neural network and the bidirectional long short-term memory network are iteratively adjusted to obtain the optimal model coefficient, further improving the three-dimensional virtual fitting effect and accuracy. Moreover, by converting two-dimensional images into three-dimensional images based on perspective transformation, not only the difficulty of obtaining three-dimensional images is effectively reduced, but also because the resolution and clarity of the image are not involved in the transformation process, the three-dimensional imaging effect is guaranteed, further improving the subsequent three-dimensional virtual fitting effect and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0037] Figure 1It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Embodiment 1:

[0040] Embodiment 1 of the present invention discloses an intelligent virtual fitting method based on three-dimensional human body modeling, such as Figure 1 As shown, including:

[0041] Step 1: Obtain two-dimensional images of the user's body and the clothes to be tried, and synthesize three-dimensional images of the user's body and the clothes to be tried based on perspective transformation adjustment, and use a three-dimensional convolutional neural network to extract three-dimensional features of the user's body and the clothes to be tried.

[0042] The three-dimensional images of the user's body and the clothing to be tried are synthesized based on perspective transformation adjustment, specifically:

[0043] The two-dimensional images of the user's body and the clothes to be tried are transformed into left and right eye images to obtain binocular parallax images to achieve stereoscopic vision, and the distance between the left and right eye images is adjusted to create a convergence angle, and the left and right eye images after the distance adjustment are synthesized to obtain a three-dimensional image of the user's body and the clothes to be tried. The present invention performs image transformation on the two-dimensional image without involving the resolution and clarity of the image, thereby ensuring that the image quality of the 3D image is the same as that of the original two-dimensional image, and ensuring the three-dimensional imaging effect.

[0044] The two-dimensional images of the user's body and the clothing to be tried are transformed into left and right eye images, specifically:

[0045] Extract the size of the two-dimensional image, scan the two-dimensional image line by line, so as to obtain the size of each line of the two-dimensional image, linearly scale the image of each line in sequence according to the principle of one end being larger and the other end being smaller, and obtain the first image, so as to reduce the amount of data processing, increase the data processing efficiency, and further realize the real-time conversion of the two-dimensional image;

[0046] The first image is mirror-processed, that is, the first image is mirror-processed along a vertical direction to obtain a second image, and the first image and the second image are used as left-eye and right-eye images.

[0047] Adjust the distance between the left and right eye images, specifically:

[0048] One of the left and right eye images is displaced horizontally so that the distance between the left and right eye images reaches a preset distance, thereby creating binocular parallax and a convergence angle. The present invention has concluded through multiple experiments that when the distance between the images changes, the corresponding convergence angle changes, that is, as the distance between the left and right eyes changes, the convergence angle will also change accordingly. Therefore, the present invention creates a convergence angle by adjusting the distance between the left and right eyes.

[0049] The 3D convolutional neural network is composed of multiple 3D feature processing units connected in sequence. Each 3D feature processing unit includes: a 3D convolution block, an activation layer, and a 3D pooling layer. Each 3D convolution block includes multiple 3D convolution kernels. The 3D convolution block is used to perform 3D convolution processing on the input image and extract 3D convolution features. The activation layer is used to add nonlinear features to the 3D convolution features. The pooling layer is used to reduce the dimensionality of the nonlinear features.

[0050] The output of the nth 3D convolution kernel in the i-th 3D feature processing unit among the multiple 3D feature processing units is as follows:

[0051]

[0052] Among them, F1 is the activation function used by the i-th three-dimensional feature processing unit; T (i-1)m (x+u, y+v, z+s) is the output of the mth 3D convolution kernel in the i-1th 3D feature processing unit;

[0053] is the sparse matrix corresponding to the nth three-dimensional convolution kernel in the i-th three-dimensional feature processing unit;

[0054] b in is the bias term of the i-th 3D feature processing unit; u, v, s are the sampling steps in the x, y, and z directions.

[0055] Step 2: Based on the three-dimensional features, a bidirectional long short-term memory network is used to predict the three-dimensional text description of the user's body and the clothes to be tried on. In combination with the three-dimensional features, a three-dimensional deconvolutional neural network is used to output a three-dimensional fitting image.

[0056] The prediction of the 3D text description based on the user's body and the clothes to be tried, combined with the 3D features, is as follows:

[0057] The three-dimensional features of the user's body and the corresponding feature matrix and feature vector in the three-dimensional text description, as well as the three-dimensional features of the clothes to be tried and the corresponding feature matrix and feature vector in the three-dimensional text description are multiplied respectively, and the feature matrices and feature vectors corresponding to the two features obtained after the multiplication are concatenated as the input of the three-dimensional deconvolutional neural network.

[0058] The three-dimensional deconvolution neural network is composed of multiple three-dimensional restoration processing units connected in sequence. Each three-dimensional restoration processing unit includes: a three-dimensional transposed convolution block, an activation layer, and an anti-pooling layer; the three-dimensional transposed convolution block is used to perform three-dimensional deconvolution processing on the input image to obtain a three-dimensional deconvolution restoration image; the activation layer is used to add nonlinear features to the three-dimensional transposed convolution; and the anti-pooling layer is used to restore features.

[0059] Step 3: Determine the authenticity of the three-dimensional fitting image based on the combination of three-dimensional convolution and fully connected layers, determine the first loss by calculating the cross entropy, and iteratively adjust the parameters of the three-dimensional convolutional neural network; locate the user body and the clothes to be tried on in the three-dimensional fitting image based on the target detection model, and use the bidirectional long short-term memory network to predict the three-dimensional text description of the user body and the clothes to be tried on after the three-dimensional fitting, calculate the second loss based on the term frequency-inverse document frequency, and iteratively adjust the parameters of the bidirectional long short-term memory network; extract the three-dimensional features of the user body and the clothes to be tried on after the three-dimensional fitting based on the three-dimensional convolutional neural network, calculate the third loss by the Euclidean distance, and iteratively adjust the parameters of the three-dimensional convolutional neural network; when the weighted sum of the first loss, the second loss, and the third loss is less than the preset loss threshold, the final intelligent virtual fitting model is obtained.

[0060] The authenticity of the 3D fitting image is determined based on the combination of 3D convolution and fully connected layers, specifically:

[0061] It consists of a plurality of three-dimensional feature restoration units connected in sequence and a fully connected layer; each three-dimensional feature restoration unit includes: a three-dimensional convolution block, an activation layer, and a discard layer; the fully connected layer is used to make authenticity judgment based on the final output of the plurality of three-dimensional feature restoration units connected in sequence, and the judgment result includes the probability that the three-dimensional fitting image is a real three-dimensional image and the probability that it is a false three-dimensional image.

[0062] Based on the authenticity of the 3D fitting image, the cross entropy is calculated to determine the first loss, as follows:

[0063] L 1 =E x [logD(x)]+E x,t [log(1-D(G(x,t)))];

[0064] Wherein, D(x) is the result of determining the authenticity of the three-dimensional fitting image; and D(G(x, t)) is the result of determining the authenticity of the three-dimensional image of the user's body.

[0065] The third loss is calculated by Euclidean distance as follows:

[0066]

[0067] Among them, H is the number of three-dimensional features; K is the dimension of the three-dimensional features; x 1k The three-dimensional features before three-dimensional fitting and the three-dimensional features after three-dimensional fitting.

[0068] The embodiment of the present invention discloses an intelligent virtual fitting method based on three-dimensional human body modeling. Based on extracting three-dimensional features of three-dimensional images at different scales through three-dimensional convolutional neural networks, the present invention introduces three-dimensional text descriptions using bidirectional long short-term memory network predictions, and performs multiplication, splicing and fusion, and jointly performs three-dimensional fitting prediction output, effectively improving the three-dimensional virtual fitting effect and accuracy. In addition, the authenticity of the three-dimensional fitting image is determined by combining three-dimensional convolution and fully connected layers, and the first loss is determined by calculating the cross entropy, and the three-dimensional text description prediction of the user's body and the clothes to be tried after three-dimensional fitting is performed based on the bidirectional long short-term memory network, and the second loss is calculated based on the term frequency-inverse document frequency, and the three-dimensional features of the user's body and the clothes to be tried after three-dimensional fitting are extracted based on the three-dimensional convolutional neural network, and the third loss is calculated by the Euclidean distance, and the three-dimensional convolutional neural network and the bidirectional long short-term memory network are iteratively adjusted to obtain the optimal model coefficient, further improving the three-dimensional virtual fitting effect and accuracy. Moreover, by converting two-dimensional images into three-dimensional images based on perspective transformation, not only the difficulty of obtaining three-dimensional images is effectively reduced, but also because the resolution and clarity of the image are not involved in the transformation process, the three-dimensional imaging effect is guaranteed, further improving the subsequent three-dimensional virtual fitting effect and accuracy.

[0069] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0070] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent virtual fitting method based on three-dimensional human body modeling, characterized in that: include: Step 1: obtaining two-dimensional images of the user's body and the clothes to be tried, and synthesizing three-dimensional images of the user's body and the clothes to be tried based on perspective transformation adjustment, and extracting three-dimensional features of the user's body and the clothes to be tried using a three-dimensional convolutional neural network; Step 2: Based on the three-dimensional features, a bidirectional long short-term memory network is used to predict the three-dimensional text description of the user's body and the clothes to be tried on, and combined with the three-dimensional features, a three-dimensional deconvolution neural network is used to output a three-dimensional fitting image; Step 3: Determine the authenticity of the three-dimensional fitting image based on the combination of three-dimensional convolution and fully connected layers, determine the first loss by calculating the cross entropy, and iteratively adjust the parameters of the three-dimensional convolutional neural network; locate the user body and the clothes to be tried on in the three-dimensional fitting image based on the target detection model, and use the bidirectional long short-term memory network to predict the three-dimensional text description of the user body and the clothes to be tried on after the three-dimensional fitting, calculate the second loss based on the term frequency-inverse document frequency, and iteratively adjust the parameters of the bidirectional long short-term memory network; extract the three-dimensional features of the user body and the clothes to be tried on after the three-dimensional fitting based on the three-dimensional convolutional neural network, calculate the third loss by the Euclidean distance, and iteratively adjust the parameters of the three-dimensional convolutional neural network; when the weighted sum of the first loss, the second loss, and the third loss is less than the preset loss threshold, the final intelligent virtual fitting model is obtained.

2. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 1, the three-dimensional image of the user's body and the clothing to be tried is synthesized based on perspective transformation adjustment, specifically: The two-dimensional images of the user's body and the clothes to be tried are perspective transformed to obtain left-eye and right-eye images, and the distance between the left-eye and right-eye images is adjusted to synthesize the left-eye and right-eye images after the distance adjustment to obtain a three-dimensional image of the user's body and the clothes to be tried.

3. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 2, characterized in that: The two-dimensional images of the user's body and the clothing to be tried are perspective transformed to obtain left and right eye images, specifically: Extracting the size of the two-dimensional image, scanning the two-dimensional image line by line, and linearly scaling the image of each line in turn to obtain a first image; Perform mirror processing on the first image to obtain a second image, and use the first image and the second image as the left-eye and right-eye images.

4. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 2, characterized in that: Adjust the distance between the left and right eye images, specifically: One of the left and right eye images is displaced in the horizontal direction so that the distance between the left and right eye images reaches a preset distance, so as to create binocular parallax and convergence angle.

5. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 1, the three-dimensional convolutional neural network is composed of a plurality of three-dimensional feature processing units connected in sequence, each of the three-dimensional feature processing units includes: a three-dimensional convolution block, an activation layer, and a three-dimensional pooling layer, and each of the three-dimensional convolution blocks includes a plurality of three-dimensional convolution kernels; the three-dimensional convolution block is used to perform three-dimensional convolution processing on the input image to extract three-dimensional convolution features; the activation layer is used to add nonlinear features to the three-dimensional convolution features; the pooling layer is used to reduce the dimensional size of the nonlinear features; The output of the nth three-dimensional convolution kernel in the i-th three-dimensional feature processing unit among the plurality of three-dimensional feature processing units is as follows: Among them, F1 is the activation function used by the i-th three-dimensional feature processing unit; T (i-1)m (x+u, y+v, z+s) is the output of the mth 3D convolution kernel in the i-1th 3D feature processing unit; is the sparse matrix corresponding to the nth 3D convolution kernel in the i-th 3D feature processing unit; b in is the bias term of the i-th 3D feature processing unit; u, v, s are the sampling steps in the x, y, and z directions.

6. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 2, based on the prediction of the three-dimensional text description of the user's body and the clothing to be tried, combined with the three-dimensional features, specifically: The three-dimensional features of the user's body and the corresponding feature matrix and feature vector in the three-dimensional text description as well as the three-dimensional features of the clothing to be tried and the corresponding feature matrix and feature vector in the three-dimensional text description are multiplied respectively, and the feature matrices and feature vectors corresponding to the two features obtained after the multiplication are concatenated as input of the three-dimensional deconvolutional neural network.

7. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 2, the three-dimensional deconvolution neural network is composed of a plurality of three-dimensional restoration processing units connected in sequence, and each of the three-dimensional restoration processing units includes: a three-dimensional transposed convolution block, an activation layer, and an anti-pooling layer; the three-dimensional transposed convolution block is used to perform three-dimensional deconvolution processing on the input image to obtain a three-dimensional deconvolution restoration image; the activation layer is used to add nonlinear features to the three-dimensional transposed convolution; and the anti-pooling layer is used to restore features.

8. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 3, the authenticity of the three-dimensional fitting image is determined based on the combination of the three-dimensional convolution and the fully connected layer, specifically: It consists of a plurality of three-dimensional feature restoration units connected in sequence and a fully connected layer; each of the three-dimensional feature restoration units includes: a three-dimensional convolution block, an activation layer, and a discard layer; the fully connected layer is used to make authenticity judgment based on the final output of the plurality of three-dimensional feature restoration units connected in sequence.

9. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 3, based on the authenticity of the three-dimensional fitting image, the cross entropy is calculated to determine the first loss, as follows: L1=E x [logD(x)]+E x,t [log(1-D(G(x,t)))]; Wherein, D(x) is the result of determining the authenticity of the three-dimensional fitting image; and D(G(x, t)) is the result of determining the authenticity of the three-dimensional image of the user's body.

10. The intelligent virtual fitting method based on three-dimensional human body modeling according to claim 1, characterized in that: In step 3, the third loss is calculated by Euclidean distance as follows: Among them, H is the number of three-dimensional features; K is the dimension of the three-dimensional features; x 1k The three-dimensional features before three-dimensional fitting and the three-dimensional features after three-dimensional fitting.

Images