Description:FORM 2
THE PATENTS ACT 1970
[39 OF 1970]
&
THE PATENTS (AMENDMENT) RULES, 2006
COMPLETE SPECIFICATION
[See Section 10; rule 13]

“DESIGNING FASHIONABLE CLOTHES WITH GENERATIVE ADVERSARIAL NETWORKS, UTILIZING STYLEGAN AND GANSPACE”

We,
1. Sujata
NIIT University, NH-8, Delhi-Jaipur Highway, Neemrana (Rajasthan)-301705, India
2. Yajur
D403, Tower D, Mahindra Aura, Sector-110A, Gurgaon, India
3. Eshanya Kamra
Belleza 43, Emaar Marbella, Sec-66, Gurgaon, India
4. Arpita Nagpal
Bharati Vidyapeeth Institute of computer application and management, Delhi

The following specification particularly describes the invention and the manner in which it is to be performed:

FIELD OF THE INVENTION

The present disclosure relates to the generative adversarial network in designing fashionable items. More particularly, the present invention relates to designing fashionable cloths.

BACKGROUND

In recent decades, deep learning has achieved significant success in numerous applications. It is becoming popular as it enables computer to learn and understand the real world through hierarchy of techniques and multiple layers, allowing machine grasp complex concepts from simple ones. One of the important applications of deep learning in fashion industry is to design and generate stylish and trendy cloths based on system’s as well as user’s choices. Dressing according to color, texture, shape, and other matching rules can significantly impact perception, making individuals appear taller or thinner and showcasing personal style. Fashion is an ever-present aspect of people's lives, although it is subject to change. People utilize fashion as a means of self-expression, achieving a distinctive style through their personal combination of clothing items. When individuals engage in online shopping for clothes, they encounter numerous options regarding styles; however, the availability of new fashion trends enhances user engagement by introducing trendy apparel choices. Recent advancements in artificial intelligence and machine learning have paved the way for fruitful collaborations between the fields of fashion and computer-based design.
Researchers have been increasingly interested in using AI to address the needs of the fashion industry, which has experienced rapid growth. The recent advancements in deep learning, coupled with powerful computational systems, have yielded impressive results particularly in tasks related to filling in missing parts of images. Notable fashion companies like Tommy Hilfiger and Macy's are applying Natural Language Processing techniques to create personalized user interfaces. Several applications have already been proposed and implemented that offer users customized recommendations based on their search history, previous purchases, and even items they've saved for later. However, there are also systems that generate new images displaying individuals wearing different outfits by utilizing posture integration and image refinement techniques. As a result, these systems do not generate new designs. Some existing systems have the capability to analyse the styles of clothing that interest the user by extracting visual features and style distribution among other characteristics. They also combine features from different types of garments such as tops and bottoms. On the other hand, some prior art utilize textures provided by users to synthesize objects that match those texture suggestions across various categories including clothing, shoes, purses, etc.
SUMMARY OF THE INVENTION:

The present invention focuses on furnishing consumers with personalized experiences by utilizing generative adversarial networks (GAN) to create and design garments that incorporate both up-to-date trends and users' preferences as inputs. GANs are deep learning models capable of generating images based on extensive datasets and can be used to generate new and unique clothing designs that align with current fashion trends. The present invention presents a novel approach using deep learning-based Generative Adversarial Networks (GAN) to generate new and trendy clothes depends on the latest trends and user choices. The present invention uses a dataset of 20000 images with 64x64 and 512x512 resolutions. First it is implemented using DCGAN as an application of deep learning to design new and innovative clothes. Further, the present invention utilizes another generative adversarial network StyleGAN as an improvement for the quality of generated cloths. The advantageous improvement of the present invention is that the novel StyleGAN presents the better performance as compared to DCGAN. Therefore, the present invention relates to an improved and efficient way of generating and presenting fashionable items i.e., cloths using StyleGan and GanSpace based on users’ choices. Detailed analysis proves the efficiency of GAN based model as a product and service in fashion industry.
The present invention provides users with a comprehensive fashion experience by offering various functionalities. The main innovation lies in the ability to design new clothes based on current fashion trends and user preferences. Although obtaining high-quality scans can be challenging due to hardware limitations, the proposed architecture successfully converts low-resolution image inputs into high-resolution image outputs. In addition to its service-based approach, the present invention also aims at establishing a strong brand presence in order to create a platform where users can easily access newly generated clothing outfits.

BRIEF AND DETAILED DESCRIPTION OF THE DRAWINGS
It should be understood at the outset that although illustrative implementations of embodiments are illustrated below, the system and method may be implemented using any number of techniques. The disclosure should in no way be limited to the illustrative implementations, drawings, and techniques illustrated below, including the exemplary design and implementation illustrated and described herein, but may be modified within the scope of the appended claims along with their full scope of equivalents.
The term “some” and “one or more” as used herein is defined as “one, or more than one, or all.” Accordingly, the terms “one,” “more than one,” but not all” or “all” would all fall under the definition of “some.” The term “some embodiments” or “one or more embodiments” may refer to one embodiment or several embodiments or all embodiments. Accordingly, the term “some embodiments” is defined as meaning “one embodiment, or more than one embodiment, or all embodiments.”
The terminology and structure employed herein are for describing, teaching, and illuminating some embodiments and their specific features and elements and do not limit, restrict, or reduce the spirit and scope of the claims or their equivalents.
More specifically, any terms used herein such as but not limited to “includes,” “comprises,” “has,” “have” and other grammatical variants thereof do not specify an exact limitation or restriction and certainly do not exclude the possible addition of one or more features or elements, unless otherwise stated, and must not be taken to exclude the possible removal of one or more of the listed features and elements, unless otherwise stated with the limiting language “must comprise” or “needs to include.”

FIG. 1 illustrates a generative adversarial network (GAN) and its working. A Generative Adversarial Network (GAN) is made up of two interconnected systems. an image is fed as input., One system of GAN, the Generator, will attempt to generate the data. The data generated will be a form of the input image for second network with some added noise. The second network of GAN called as Discriminator, distinguishes the generated image between real and fake. By fooling the discriminator about half of the images, it tries to make the discriminator detect the fake images as real ones. GAN is most widely used in image translation, wherein, the images are faked in terms of changing seasons or surroundings. The error produced by the discriminator will be used to improve the efficiency of generation of images by the generator.

The present invention describes a novel approach following two steps, the GAN and the clothes detector. In the present invention processes dataset which consists of various fashion items ranging from foot wear (shoes, flip flops, sandals) to apparel section (tops, t-shirts, shirts, bottom wear, dresses, etc). The present invention have trained the model on the apparel section which consists of about 20,000 images (tops, t-shirts, shirts, bottom wear, dresses, etc.).
The method of present invention consists feeding the test images as input training data to custom DCGAN or Deep-Convolutional Generative Adversarial Network, which consists of a Discriminator and a Generator. Generator and Discriminator play a two-player minimax game which can be given with the following equation.

minG maxD V (D, G) = Ex~pdata(x) [log(D(x))] + Ez~pz(z) [log(1 - D(G(z)))] (1)

where D(x) tells us whether the image is fake or real in probabilistic terms, G(z) is generated output from a given z. Pdata and Pz are the statistical distribution of the real and fake data.

RESULT AND DISCUSSIONS
Firstly the present invention discloses the uses of 64x64 images for the phase 1 of the generator/discriminator training. The architecture of generator and discriminator can be seen in Figures-2 and Figure-3.

Figure 2. Architecture of the Generator


Figure 3. Architecture of the Discriminator

Binary Cross Entropy loss criteria has been used to evaluate the losses of the GAN model and to back propagate the gradients. ADAM optimizer is used in both of the GAN’s components. The customized framework is implemented in Pytorch.

The whole setup is run for 100 epochs on Apple M1 (8 Cores CPU and 8 Cores GPU) with 8GB RAM. The losses can be seen in the Figure-4 and A. The generated clothing set can be visualized through a series of grids of images over a period of 100 epochs from Figs. Aa through Af.

Figure. 4 The Generator and discriminator loss throughout the 100 epochs.

Figure. A Training Output


Figure Aa. Generated Images at Epoch 0
Figure Ab. Generated Images at Epoch 18

Figure Ac. Generated Images at Epoch 32
Figure Ad. Generated Images at Epoch 63
Figure Ae. Generated Images at Epoch 84
Figure Af. Generated Images at Epoch 100

StyleGAN on 512x512 images

To train the model more efficiently than DCGAN, proposed research used StyleGAN [6] to train the images for 512x512 resolution. StyleGAN works on a concept of progressive gains and bilinear sampling. Progressive gains makes training better by taking in low resolution images and then upsampling it to higher resolution images. Fig.7 shows the architecture of StyleGAN that has 18 convolutional layers where each upsampling layer is followed by 2 convolution layers. It starts at 4x4x512 and goes up to 1024x1024x512.
StyleGAN allows us to have more control over the variation in the characteristics of the generated image. First, a latent vector z is connected into a fully connected network of 8 layers whose output shape (W) is the same as the input shape (512x1). Next AdaIN (Adaptive Instance Normalization) embeds styling into the generated image at each layer by passing the W modified by some learned affine transformations (A).

Figure. 7 StyleGAN architecture (Liu, et al. (2017))

4.2 Typing and Font Specification
GANSpace enables inspecting and manipulating of GANs. It allows users to control the directions in GAN’s latent space, which can help understand how the model works and generate new outputs. GANSpace works by taking an image generated by a GAN and breaking it down into its latent space representation. The latent space is the space of all possible input values that can be used to generate an image using the GAN. GANSpace then generates a navigable 3D space where users can explore and manipulate the latent space to create new images.
Some of the features of GANspace include the ability to exploratively traverse the latent space, to search for features in the latent space and to create new images using latent directions. It can also modify existing images by changing certain features in the latent space.
GANSpace uses Principal Component Analysis(PCA) to find direction in an isotropic gaussian distribution. It is an unsupervised method for learning latent direction. These principal components capture the most significant variations in the generated images. GANSpace then allows users to manipulate these principal components, adjusting the features of the generated images. For example, users would be able to adjust color, sleeve length, necklines, structure and style of the clothes.

GANSpace involves applying PCA in the latent space for StyleGAN and feature space for BigGAN. BigGAN can also be modified to StyleGAN to allow layer-wise style mixing without retraining. The most basic representation consists of a probability distribution from which a latent vector is sampled, and a neural network that produces an output image. This network can be decomposed into a series of intermediate layers. In the BigGAN model, Skip-z inputs are used in the intermediate layers, along with a class vector as input. In the StyleGAN model, a constant input is used for the first layer, and the output is controlled by a nonlinear function of z in the intermediate layers. Allowing each layer to have its own style vector (wi) enables powerful "style mixing" in generating images.

GANSpace believes that the most variations can be captured in the early layers of the model. For StyleGAN, principal axes of p(w) can be identified by computing PCA of the M(zi) values corresponding to N random vectors z1:N. The resulting basis V for W can be used to edit a new image by varying the PCA coordinates x. For BigGAN, which lacks a learned z distribution and w latent space, PCA is computed at an intermediate layer i and transferred to the z latent space using linear regression. The resulting principal directions can be used to edit new images by adding offsets along the columns of the principal direction matrix.

The Figure. 8 shows principal components from which authors selected certain characteristics of the cloths. For example, the 0th component (C0) can be labelled as sleeve length, C2 can be labelled as coats, C4 and C5 can be labelled as increase in the dark tone of the cloth and C6 can be different necklines.
Figure.9a. Shows combinations of two top wears, a jacket and a sweatshirt. The output has the structure of a jacket but the simplicity and color of the sweatshirt.
Figure.9c. shows a combination of 2 tops into another top that has the color in the similar tone of top 1 but a slightly darker shade as top 2 is black. Changing top 2 with another black top (Figure. 9d.) gave a nice look of a grey coat. Altering dimensions like less cleavage and darkness of the color modified the dress to become pink from grey and opened up more neckline which can be seen in Figure.9e.

Figure 8. The first 14 principal components to find the respective characteristics of the cloth.

Figure. 9a. Combination of two top wears.

Figure. 9b. Combination of a t-shirt and a coat into a different female dress

Figure. 9c. Combination of a top and a sweatshirt.

Figure. 9d. Combination of two tops into a coat.

Figure. 9e. An even more modified version

The present invention discloses steps in providing an efficient and economical solution for the clothing industry for designing new and innovative clothes. Resulted patterns and designs have a touch of both old and new fashion clothes. Though the results are good, but still there is a scope of training our model for longer iterations and larger datasets on more powerful machines using better hyperparameter tuning to get the improved version of the proposed model. This work took topwear as our primary category for training models, we would like to expand by going into more clothing domains. For now, it is just trained on a dataset which contains 9000+ apparel images with clean backgrounds.
Although its contents, architecture and working methodology proved to be a challenge for the authors still we aim to enhance it for a larger dataset. Major issues in training the model are due to its high RAM and GPU requirements. Each training took more than 10 hours to complete.

, Claims:We Claim:

1. An autonomous GAN based system that comprises the following:
a) a generator, attempts to generate the data in the form of some input images with random noises,
b) a discriminator, tries to distinguish the generated image between real and fake
c) a preprocessing unit to smoothen the input by improving data discretization.

2. The system as claimed in claim 1, wherein an improved method for generating or designing the new cloths as per client’s and business’s recommendation,

3. The system as claimed in claim 1, which further comprising model which is executable by the processor, wherein the model comprising:

• Cloths Recommendation by Product recommendation system
• Managing by improved security module
• Authentication by Inventory and account management module.

Dated this 20th August of 2024

[ABHISHEK SHARMA]
Digitally signed
PATENT AGENT NUMBER: IN/PA 2641
OF LexAniv
AGENT FOR THE APPLICANTS