A complete clothing generation model based on clothing matching technology and its implementation method

Abstract

The present invention discloses a complete clothing generation model based on clothing matching technology and its implementation method. The model includes a suit generator, a suit discriminator, and a matching supervision module. The suit generator includes multiple clothing item generators, which convert input clothing images and target clothing outline masks into corresponding matching clothing suit images. Each clothing item generator is responsible for generating an image of the corresponding target clothing item. The suit discriminator includes multiple clothing item discriminators, each of which is responsible for ensuring the authenticity of the generated clothing item image. The matching supervision module includes a convolutional neural network feature extractor and a bidirectional long short-term memory network, which are used to enhance the matching degree of the generated clothing suit and improve the matching degree of the generated clothing with the given clothing. The clothing item generator must meet the identification requirements of the corresponding clothing item discriminator.

Description

Technical Field

[0001] The present invention belongs to the field of generative models and clothing matching, and in particular relates to a complete clothing generation model for matching based on a generative model and an implementation method thereof. Background Art

[0002] Fashion, with its ability to express individuality and shape culture, plays an increasingly important role in today's society, especially in the current internet environment. In recent years, with the rapid adoption of the internet and e-commerce, the online shopping economy has experienced rapid growth. In particular, online apparel purchases and the online clothing supply chain have shown a rapid year-on-year growth trend, creating significant business opportunities for clothing recommendation methods. According to a report, "In recent years, the global apparel sales market has continued to grow rapidly, with the global related market maintaining an annual growth rate of 7.2%. Traditional offline sales have begun to decline. As users migrate, retailers are gradually shifting their traditional offline sales budgets to online e-commerce businesses and increasing their efforts in customer recommendations.""According to the survey, the public literature shows ([1] McAuley, Julian, et al. "Image-based recommendations on styles and substitutes." Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 2015; Han, Xintong, et al. "Learning fashion compatibility with bidirectional lstms." Proceedings of the 25th ACM international conference on Multimedia.2017; [2] Vasileva, Mariya I., et al. "Learning type-aware embeddings for fashion compatibility." Proceedings of the European Conference on Computer Vision (ECCV). 2018; [3] Cui, Zeyu, et al. "Dressingas a whole: Outfit compatibility learning based on node-wise graph neural networks." The World Wide Web Conference.2019; [4] Li, Xingchen, et al. "Hierarchical fashion graph network for personalized outfit Currently, the most commonly used clothing matching methods mainly use discriminant models. That is, given a set of clothing, the clothing matching model evaluates it to determine whether it matches. These discriminant models mainly use metric learning and graph representation learning to model the clothing matching problem. However, these models can only evaluate whether existing clothing items match and cannot generate new matching clothing sets.In addition to the research on clothing matching methods mentioned above, there are also a small number of studies focusing on the field of clothing matching. In the literature published by Liu et al. (Liu, Linlin, et al. "Collocating clothes with generative adversarial networks cosupervised by categories and attributes: a multidiscriminator framework." IEEE transactions on neural networks and learning systems 31.9 (2019): 3540-3554.), a technology for generating bottoms from a given top based on a generative model was introduced. However, this technology has four limitations: (1) it is limited to the task of one-to-one (top and bottom) generation model; (2) the generation process lacks interpretability; (3) it does not utilize the relationship between the elements or styles of the matching between clothing; and (4) the generation process is uncontrollable and cannot generate clothing with the user's desired outline. A subsequent paper published by Yu et al. (Yu, Cong, et al. "Personalized fashion design." Proceedings of the IEEE / CVF International Conference on Computer Vision, 2019) also focused on solving the problem of generating tops and bottoms, and similarly encountered the four aforementioned issues. In many practical applications, people need to target a specific clothing item and hope to complete the remaining clothing to construct a complete set of matching outfits. Sometimes, they also want to specify the outline of the generated clothing. However, existing models can only evaluate the matching degree of a given clothing set or convert between tops and bottoms.

[0003] In recent years, the rapid development of generative adversarial network technology in visual fields such as images and videos has led to a large number of applications based on generative models emerging in the market. Among them, the great success of image translation in image-to-image conversion tasks has made it possible to convert generative adversarial network models into applications in real-world scenarios. In this patent, relying on existing generative adversarial network technology, especially image translation learning algorithms, existing clothing and potential matching clothing are closely connected to improve the feasibility of recommendations and matching in online shopping, and to make clothing recommendation tasks more accurate and natural for users. This patent designs a whole set of clothing generation model using clothing matching technology. First of all, the design inspiration of this patent comes from the current whole set of matching clothing. Each piece of clothing has matching elements that enhance the overall style. In addition, the clothing generation task should not only generate opinions, but should generate multiple pieces of clothing at one time, which can be matched with existing clothing to form a complete set of clothing. Summary of the Invention

[0004] The purpose of the present invention is to provide a model for generating a complete set of clothing based on clothing matching technology, aiming to solve the above-mentioned technical problems, namely: (1) Most of the existing clothing matching technologies can only evaluate the matching degree of a given clothing set; (2) A small number of technologies for generating matching clothing can only complete the conversion between upper and lower garments and cannot be extended to generate a complete set of matching clothing; (3) The generation process lacks corresponding explainability; and (4) The generation process is difficult for users to control, which may produce clothing that users do not want. The purpose of the present invention is achieved through the following measures:

[0005] A complete clothing generation model based on clothing matching technology includes a suit generator, a suit discriminator, and a matching supervision module. The suit generator includes multiple clothing item generators, and the suit discriminator includes multiple clothing item discriminators. Each clothing item generator is responsible for generating an image of a corresponding target clothing item, and each clothing item discriminator is responsible for ensuring the authenticity of the generated clothing item image. The matching supervision module includes a convolutional neural network feature extractor and a bidirectional long short-term memory network, which are used to enhance the matching degree of the generated clothing suit and improve the matching degree between the generated clothing and the given clothing. In addition, the clothing item generator must not only meet the identification requirements of the corresponding clothing item discriminator, but also generate a synthetic image that is close to the real image in terms of similarity distance; preferably, the similarity distance is the L1 distance between the generated image and the real image, as well as the perceptual loss distance between the generated image and the real image.

[0006] Furthermore, the clothing item generator adopts an "encoder-semantic alignment module-decoder" structure; the clothing item discriminator adopts a multi-scale discriminator; the matching supervision module adopts the feature extraction part of a classification model pre-trained on image net as a feature extractor, and a bidirectional long short-term memory network models the matching of the entire set of clothing. The bidirectional long short-term memory network adopts a cross-entropy loss function as its objective function.

[0007] Furthermore, the encoder adopts a stack of multi-layer convolutional neural networks, and the decoder adopts an alternating stack of multi-layer upsampling layers and convolutional neural networks. Each of the semantic alignment modules includes two convolutional neural networks, which are respectively used to extract the high-level features of the given clothing after passing through the encoder and the high-level features of the contour mask; the semantic alignment module also includes a correspondence calculation layer for calculating the correspondence matrix between the above two high-level features, and this matrix contains the correspondence size between any positions between the above two high-level features; the semantic alignment module also includes a feature alignment layer for aligning the features obtained by the encoder according to the above-mentioned correspondence matrix.

[0008] Furthermore, when training the clothing item discriminator, the loss function of the image generated by the clothing item generator should be maximized, and the loss function of the real image should be minimized; when training the clothing item generator, the image generated by the generator should tend to be the real image, minimizing the corresponding clothing item discriminator loss.

[0009] The implementation method of the entire clothing generation model based on clothing matching technology includes the following steps:

[0010] A. Constructing a clothing dataset: Constructing a clothing dataset containing image information, contour masks of all images, and information about all image categories;

[0011] B. Pre-trained Models: The entire clothing generation model based on clothing matching technology requires pre-training two models: a VGG model for perceptual loss and a matching supervision module. The matching supervision module includes a convolutional neural network feature extractor and a bidirectional long short-term memory network. The VGG model for perceptual loss is pre-trained on ImageNet and then fine-tuned for attribute prediction on public clothing datasets such as DeepFashion. The fine-tuned model is then used for perceptual loss calculation. The matching supervision module is trained for matching prediction on the clothing matching dataset described in A.

[0012] C. Design a complete clothing generation model based on clothing matching technology: the complete clothing generation model includes a suit generator, a suit discriminator and the matching supervision module described in B. The suit generator contains multiple clothing item generators, and the suit generator completes the conversion of the input clothing image and the target clothing outline mask into the corresponding matching clothing suit image; the suit discriminator contains multiple clothing item discriminators, each clothing item discriminator is responsible for determining the authenticity of the generated clothing item and supervising the corresponding clothing item generator; the matching supervision module is used to enhance the matching degree of the generated clothing suit and improve the matching degree between the generated clothing and the given clothing.

[0013] Furthermore, the step A comprises the following steps:

[0014] A1. Build a clothing matching dataset, which includes clothing images with matching information, category annotations, and corresponding image contour masks. Each outfit also includes the user's preference level.

[0015] A2. For the clothing matching dataset A1, perform data cleaning on it, select the K most popular clothing sets based on their popularity, and then label the outline and category of each clothing item. Finally, the selected clothing sets and their corresponding outline masks are used as the dataset.

[0016] Furthermore, the step B comprises the following steps:

[0017] B1. The VGG model for the perceptual loss contains 13 convolutional layers, with a specific structure of "two convolutional layers - max pooling layer - two convolutional layers - max pooling layer - three convolutional layers - max pooling layer - three convolutional layers - max pooling layer - three convolutional layers - max pooling layer." The input image is first resampled to 224×224 and passed through the VGG model for the perceptual loss to obtain a 224×224 feature map. The VGG model for the perceptual loss calculates the L1 distance between the features of the generated image and the features of the real image. Through this constraint, the distance between the generated image and the real image is reduced in the dimension of high-level features.

[0018] B2. The collocation supervision module consists of two parts: a feature extractor and a bidirectional long short-term memory network. The feature extractor consists of a multi-layer convolutional neural network followed by a multi-layer residual network. The multiple images generated by the feature extractor are all passed through the feature extractor to obtain feature maps, which are then reshaped into vector form. These vectors are then input into the bidirectional long short-term memory network to model the collocation, predicting the vector form of unknown clothing given known clothing. Here, the cross-entropy loss function is used. This collocation supervision module uses the collocation clothing dataset described in step A.

[0019] B3. The VGG model corresponding to the perceptual loss and the matching supervision module keep the parameters fixed during the training of the entire clothing generation model.

[0020] Furthermore, the step C comprises the following steps:

[0021] C1. Input the complete set of matching clothing images and the contour mask of the image to be generated as supervision information into the complete clothing generation model;

[0022] C2. Adopt the adversarial training strategy for iterative training. First, fix all the parameters of the suit discriminator, perform forward propagation on the suit generator, then calculate its loss function and perform gradient descent, and update the parameters of the suit generator. In the subsequent process, fix the various parameters of the suit generator, allow the parameters of the suit discriminator to learn, perform forward propagation, then calculate its loss function and perform gradient descent, and update the parameters of the suit discriminator.

[0023] Specifically, the input to the entire clothing generation model is divided into two parts: one is a collection of clothing sets containing matching information, each of which contains multiple matching clothing items; the other is the annotated clothing contour mask information. During a single iteration, the input image and the generated image form a set of real matching images, which are fed into a pre-trained matching supervision module to supervise the matching degree of the generated clothing items. In addition, the set discriminator also supervises the authenticity of the generated clothing items. The set generator includes multiple clothing item generators, each of which consists of an "encoder-semantic alignment module-decoder." The semantic alignment module includes a two-way convolutional neural network to extract features from a given clothing image and contour mask; a correspondence calculation layer to calculate the correspondence matrix between the two features. This matrix contains the correspondence size between any positions of the two high-level features; and a feature alignment layer to align the features obtained by the encoder according to the correspondence matrix.

[0024] The present invention proposes for the first time a complete clothing generation model that can generate multiple clothing items at one time to form a suit with a given clothing item, providing a reference for users' dressing and clothing designers' clothing design. Specifically, the present invention proposes a complete clothing generation model based on clothing matching technology to generate a complete set of matching clothing images. Under the framework of the complete clothing generation model, the suit generator will generate multiple clothing items that match the input clothing image to form a matching clothing suit. The semantic alignment module in the suit generator can also play a role in generating more matching styles and relationships between matching elements. The clothing identifier plays a role in supervising the generator to generate real images. The matching supervision module supervises the matching degree of the generated suit clothing, so that the generated image is more realistic. The given contour mask can also control the generation process to generate the clothing suit required by the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the implementation method of the entire clothing generation model based on clothing matching technology of the present invention;

[0026] Figure 2 This is a structural diagram of a complete clothing generation model based on clothing matching technology of the present invention;

[0027] Figure 3 It is the loss function mechanism of the whole set of clothing generation model based on clothing matching technology of the present invention;

[0028] Figure 4 This is a result diagram of the entire clothing generation model based on the clothing matching technology of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] Attachment Figure 1 The flowchart of the whole clothing generation model based on clothing matching technology provided by the present invention is shown. Figure 2 This is a structural diagram of the entire clothing generation model based on clothing matching technology of the present invention. It is described in detail as follows.

[0031] Step S1: Constructing a clothing matching dataset. The training dataset used in this invention comes from the Polyvore.com website. Users can upload, share, and modify their own clothing matching collections, while other users can also comment on and rate others' matching collections. This invention crawls matching combinations that include image information, category information, and user preference information.

[0032] On this basis, we cleaned the collocation data crawled above, and only retained the clothing sets that included four complete categories: tops, bags, bottoms, and shoes. On this basis, we annotated the contour mask of each image obtained, and screened out the 20,000 most popular clothing sets as the data set, and then divided the data set into a training set (14,000, 70%), a validation set (2,000, 10%), and a test set (4,000, 20%).

[0033] Step S2: Pre-training model. The present invention mainly designs a pre-trained matching supervision module to guide the generation of the matching degree of a whole set of clothing, and also trains another VGG network to obtain the perceptual loss. The specific description is as follows: the VGG model corresponding to the perceptual loss is pre-trained on image net, and then the attribute prediction is performed on the public dataset DeepFashion. The fine-tuned model is used to calculate the perceptual loss; the matching supervision module described here is trained for matching prediction on the clothing dataset obtained in step S1. Specifically, this matching supervision module adopts the cross entropy loss function As its objective function:

[0034]

[0035] where f i It represents the features of the i-th clothing item obtained after the convolutional neural network. f i The implicit features obtained by the forward long short-term memory network and the backward long short-term memory network. N is the number of clothing items contained in a set of clothing, and the feature f i Then send it into the bidirectional long and short time memory network to get and Here exp is the natural exponential function, and It contains the features f of all functions in the current batch i During the pre-training process of the outfit supervision module, the outfit matching dataset constructed in step S1 is used to perform pre-training by minimizing the aforementioned loss. The parameters of this pre-trained outfit supervision module remain unchanged during the training of the entire outfit generation model, and it is used to supervise the entire outfit generation model. Figure 3 It is the loss function mechanism of the entire clothing generation model based on clothing matching technology of the present invention.

[0036] Step S3: Build a complete outfit generation model based on collocation learning. This model consists of three parts: a clothing set generator, a clothing set discriminator, and a collocation supervision module. The input of the complete outfit generation model is the complete clothing set and the corresponding contour mask of each clothing item.

[0037] Among them, the clothing set generator contains N-1 clothing item generators, where N is the number of clothing items in the whole set of clothing. The structure of each clothing item generator is "encoder-semantic alignment module-decoder", in which the encoder uses a multi-layer convolutional neural network stack, the decoder uses an alternating stack of multi-layer upsampling layers and convolutional neural networks, and the semantic alignment module contains two convolutional neural networks, a correspondence calculation layer and a feature alignment layer. The above two convolutional neural networks are respectively stacked with 4 layers of convolutional layers, which are used to extract the feature F of a given clothing item. x and the contour mask feature F y The corresponding relationship calculation layer involves the following calculations:

[0038]

[0039] in Represents the single-channel convolutional neural network in the semantic alignment module, T represents the transpose operation of the matrix, and a given clothing item passes through the i-th encoder Enc i The extracted features, is the feature extracted by the contour mask after passing through another convolutional neural network in the semantic alignment module. Here, the relationship calculation layer in the i-th corresponding semantic alignment module is used as an example. ||·|| represents the modulus of the two vectors. Through the above calculation, the corresponding relationship matrix between any two positions u and v is obtained.

[0040] Furthermore, we continue to use the relationship response matrix obtained above The feature F obtained by the i-th encoder src Perform alignment. The feature alignment operation of the feature alignment layer is:

[0041]

[0042] in Represents the features after alignment, is the normalized exponential function in the v dimension, the feature after alignment Then it is sent to the i-th decoder for feature decoding to obtain the final i-th clothing item, the final N-1 clothing item images and the given clothing item image To form a complete set of matching clothing sets.

[0043] In addition, our clothing set discriminator also contains multiple clothing item discriminators. Each clothing item discriminator adopts the architecture of multi-scale discriminator. Here, the i-th clothing item discriminator D i As an example, the discriminator D iThe loss function for distinguishing true and false samples is as follows:

[0044]

[0045] Specifically, for each image of scale s, the loss function is as follows:

[0046]

[0047] in, It means the 2 before downsampling the image x -s times, p data is the distribution of real data, Mask i Represents the contour mask of the i-th clothing item, and the minimum mean square loss function is used here.

[0048] Furthermore, the training process of the clothing set generator involves the superposition of multiple loss functions, and its overall loss function can be expressed as:

[0049]

[0050] in Represents the overall loss function of the generator. Specifically, the overall loss function consists of four parts: generating adversarial loss function L1 loss function Perceptual loss function and the paired supervision loss function where λ1 and λ2 are balance coefficients.

[0051] Furthermore, the clothing set generator can be written as follows regarding the generation of adversarial functions:

[0052]

[0053] in, is the adversarial loss function of the i-th clothing item generator, which can supervise the entire clothing generator to generate realistic images. In addition, the L1 loss function is defined as follows:

[0054]

[0055] in represents the synthesized image of the i-th clothing item, represents the real image of the i-th clothing item. In addition to the L1 loss function, the present invention also uses a perceptual loss function, which can be defined as:

[0056]

[0057] where l∈{relu1_2,relu2_2,relu3_3,relu4_3}, is the corresponding convolutional layer in the VGG network, φ l It represents the convolution function of convolution layer l.

[0058] Step S4: Train a whole outfit generation model based on matching learning. Taking a given outfit image as input and the outline masks of several other objects as reference information, the whole outfit generator generates multiple clothing items that can be matched with the given clothing item. The whole outfit discriminator continuously learns to distinguish between real and synthetic images. The whole outfit generation model training process uses the standard GAN model training process, with the outfit generator and outfit discriminator alternately performing gradient descent. Mini-batch stochastic gradient descent and the Adam algorithm are used for training.

[0059] Step S5: Given a new clothing item image and a target contour mask, the generator generates a complete set of matching clothing sets. After the training of the complete clothing set generation model is completed, the network parameters of the complete clothing set generator are obtained. The given single clothing item image and the target contour mask are input into the suit generator to generate multiple clothing items that match it. The example of the matching clothing image generated by the method of the present invention is as follows: Figure 4 shown.

Claims

1. A method for implementing a complete clothing generation model based on clothing matching technology, characterized in that: The implementation method comprises the following steps: A. Constructing a clothing dataset: Constructing a clothing dataset containing image information, contour masks of all images, and information about all image categories; B. Pre-trained Model: The entire clothing generation model based on clothing matching technology requires pre-training two models, including a VGG model for perceptual loss and a matching supervision module. The matching supervision module contains a convolutional neural network feature extractor and a bidirectional long short-term memory network. The VGG model for perceptual loss is pre-trained on ImageNet, and then attribute prediction is performed on a public clothing dataset. The fine-tuned model is used for perceptual loss calculation. The matching supervision module is trained for matching prediction on the clothing matching dataset described in A. C. Design a complete clothing generation model based on clothing matching technology: the complete clothing generation model includes a suit generator, a suit discriminator, and the matching supervision module described in B. The suit generator includes multiple clothing item generators, each of which is responsible for generating an image of a corresponding target clothing item. The suit generator completes the conversion of the input clothing image and the target clothing outline mask into the corresponding matching clothing suit image. The suit discriminator includes multiple clothing item discriminators, each of which is responsible for ensuring the authenticity of the generated clothing item image. The matching supervision module is used to enhance the matching degree of the generated clothing suit and improve the matching degree of the generated clothing with the given clothing. The clothing item generator must not only meet the identification requirements of the corresponding clothing item discriminator, but also generate a synthetic image that is close to the real image in terms of similarity distance. The clothing item generator uses an "encoder-semantic alignment module-decoder" structure; the clothing item discriminator uses a multi-scale discriminator; the matching supervision module uses the feature extraction part of a classification model pre-trained on ImageNet as a feature extractor, and a bidirectional long short-term memory network to model the matching of the entire set of clothing. The bidirectional long short-term memory network uses the cross-entropy loss function as its objective function. The encoder adopts a multi-layer convolutional neural network stack, and the decoder adopts an alternating stack of multi-layer upsampling layers and convolutional neural networks. Each of the semantic alignment modules includes two convolutional neural networks, which are respectively used to extract the high-level features of the given clothing after passing through the encoder and the high-level features of the contour mask; the semantic alignment module also includes a correspondence calculation layer for calculating the correspondence matrix between the above two high-level features. This matrix contains the correspondence size between any positions between the above two high-level features and a feature alignment layer for aligning the features obtained by the encoder according to the correspondence matrix obtained above.

2. The implementation method according to claim 1, characterized in that: The step A comprises: A1. Build a clothing matching dataset, which includes clothing images with matching information, category annotations, and corresponding image contour masks. Each outfit also includes the user's preference level. A2: Clean the data set of A1 and select the K most popular clothing sets based on their popularity. Then, label the outlines and categories of each clothing item. The selected clothing sets and their corresponding outline masks are used as the dataset.

3. The implementation method according to claim 1, characterized in that: The step B comprises: B1. The VGG model corresponding to the perceptual loss contains 13 convolutional layers, with a specific structure of "two convolutional layers - max pooling layers - two convolutional layers - max pooling layers - three convolutional layers - max pooling layers - three convolutional layers - max pooling layers - three convolutional layers - max pooling layers". The input image is first resampled to a size of 224×224, and then passed through the VGG model corresponding to the perceptual loss to obtain a feature map of size 224×224. The VGG model corresponding to the perceptual loss calculates the L1 distance between the features of the generated image and the features of the target image. Through such constraints, the distance between the generated image and the target image is reduced in the dimension of high-level features. B2. The collocation supervision module includes two parts: a feature extractor and a bidirectional long short-term memory network. The feature extractor includes a multi-layer convolutional neural network, followed by a multi-layer residual network. The multiple images generated are all passed through the feature extractor to obtain feature maps, which are then reshaped into vector form. The above vectors are then input into the bidirectional long short-term memory network to model the collocation and predict the vector form of unknown clothing given known clothing. The cross-entropy loss function is used here. This collocation supervision module uses the collocation clothing dataset described in step A. B3. The VGG model corresponding to the perceptual loss and the matching supervision module keep the parameters fixed during the training of the entire clothing generation model.

4. The implementation method according to claim 1, wherein: The step C comprises: C1. Input the complete set of matching clothing images and the contour mask of the image to be generated as supervision information into the complete clothing generation model; C2. Adopt the adversarial training strategy for iterative training. First, fix all the parameters of the suit discriminator, perform forward propagation on the suit generator, then calculate its loss function and perform gradient descent, and update the parameters of the suit generator. In the subsequent process, fix the various parameters of the suit generator, allow the parameters of the suit discriminator to learn, perform forward propagation, then calculate its loss function and perform gradient descent, and update the parameters of the suit discriminator.

5. The implementation method according to claim 1, wherein: In step C, a standard generative adversarial network model training process is adopted, the set generator and the set discriminator perform gradient descent alternately, and use small-batch stochastic gradient descent and the Adam algorithm to optimize parameters.

6. The implementation method according to any one of claims 1 to 5, characterized in that: The implementation method further includes inputting a given single clothing image and a target contour mask into a suit generator to generate multiple clothing images that match the given single clothing image.

7. The implementation method according to any one of claims 1 to 5, characterized in that: The similarity distance is the L1 distance between the generated image and the real image and the perceptual loss distance between the generated image and the real image.

8. The implementation method according to any one of claims 1 to 5, characterized in that: When training the clothing item discriminator, the loss function of the image generated by the clothing item generator should be maximized, and the loss function of the real image should be minimized; when training the clothing item generator, the image generated by the generator should be inclined to the real image, minimizing the corresponding clothing item discriminator loss.

Images