# US 2023/0410390 A1 · Method and system for generating color variants for fashion apparels

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0410390 A1
- Document ID: US20230410390A1
- Status: Published application
- Kind: A1
- Application number: 18/336,649
- Filed: 2023-06-16
- Published: 2023-12-21
- Earliest claimed: 2022-06-17
- Assignees: Tata Consultancy Services Limited
- Applicants: Tata Consultancy Services Limited
- Inventors: BAGYA LAKSHMI VASUDEVAN; KALYAN PRAKASH BAISHYA; KURUVILLA ABRAHAM; JAYAVARDHANA RAMA GUBBI LAKSHMINARASIMHA; GAURAB BHATTACHARYA; NIKHIL KILARI
- Web page: https://patentgrep.com/patent/US20230410390A1

## CPC

- G06T 11/10: Texturing; Colouring; Generation of textures or colours (retouching, inpainting or scratch removal G06T5/77)
- G06T 7/194: involving foreground-background segmentation
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL

## Abstract

State of art techniques for color regeneration are complex and fail to provide color control. Embodiments of the present disclosure provide a method and system for generating color variants for fashion apparels by providing a Fashion Apparel Regeneration-Generative Adversarial Network (FAR-GAN) to generate color variants of the fashion apparels. The FAR-GAN utilizes a two-step encoding process to encapsulate both an input image and an edge-map information along with a target color embedding branch to manipulate the color information present in the fashion apparel present in the input image that is to be changed to a desired target color. Furthermore, the color and structural information is disentangled by controlling them using a color consistency loss. The FAR-GAN can be trained end-to-end without incorporating complex multi-step process.

## Claims

### Claim 1

A processor implemented method for generating color variants for fashion apparels, the method comprising:
receiving, via one or more hardware processors, (a) an input image comprising a fashion apparel, and (b) a target color for generating a color variant of the fashion apparel, wherein the target color is represented in terms of Red Blue Green (RGB) components;
generating, via a two-step encoder of a Fashion Apparel Regeneration-Generative Adversarial Network (FAR-GAN) executed by the one or more hardware processors, a target style embedding of the fashion apparel by processing the input image, the two-step encoder comprising (a) a first level encoder generating a concatenated feature embedding representing a plurality of primary low-level features and an edge map of the fashion apparel and (b) a second level encoder generating the target style embedding from the concatenated feature embedding;
generating, via the one or more hardware processors, a target color embedding from the RGB components of the target color via multi-layer Fully Connected (FC) layers of the FAR-GAN; and
generating, via a decoder of the FAR-GAN executed by the one or more hardware processors, a synthetic target image comprising the color variant of the fashion apparel, wherein the decoder dictates the target style embedding by the target color embedding using an Adaptive Instance Normalization (AdaIN) and utilizes un-correlated Gaussian noise added to each of the plurality of decoder blocks to impose variability in a color and a style of the fashion apparel in the synthetic target image enabling the synthetic target image to appear realistic when displayed on a display screen.

### Claim 2 (depends on 1)

The method of claim 1, further comprising a discriminator connected to the decoder to determine whether the synthetic target image is a realistic or a non-realistic image by comparing the synthetic target image with the input image, wherein the FAR-GAN is trained on an adversarial loss to continuously enhance the synthetic target image.

### Claim 3 (depends on 1)

The method of claim 1, further comprising use of a Color Consistency Loss (CCL) to explicitly control color information and penalize the FAR-GAN until a color of the synthetic target image is similar to the target color, wherein the CCL utilizes a cost function L2 which is a function of the target color and a dominant color in the synthetic target image.

### Claim 4 (depends on 1)

The method of claim 1, wherein generating the target style embedding comprising:
extracting the plurality of primary low-level features of the fashion apparel via a first encoder of the first level encoder to generate an image feature embedding;
generating the edge map of the fashion apparel via a pre-trained Cycle GAN;
generating, via a second encoder of the first level encoder, an edge feature embedding from the edge map;
concatenating the image feature embedding and the edge feature embedding to generate the concatenated feature embedding; and
generating the target style embedding of the fashion apparel by processing the concatenated feature embedding via the second level encoder.

### Claim 5 (depends on 1)

The method of claim 1, wherein if the input image comprises a cluttered background, a segmentation process is applied on the input image to detect the fashion apparel from the input image prior to processing the input image by the first level encoder.

### Claim 6

A system for generating color variants for fashion apparels, the system comprising:
a memory storing instructions;
one or more Input/Output (I/O) interfaces; and
one or more hardware processors coupled to the memory via the one or more I/O interfaces,
the memory comprising a Fashion Apparel Regeneration-Generative Adversarial Network (FAR-GAN) executed by the one or more hardware processors is configured by the instructions to:
receive (a) an input image comprising a fashion apparel, and (b) a target color for generating a color variant of the fashion apparel, wherein the target color is represented in terms of Red Blue Green (RGB) components;
generate via a two-step encoder of the FAR-GAN, a target style embedding of the fashion apparel by processing the input image, the two-step encoder comprising (a) a first level encoder generating a concatenated feature embedding representing a plurality of primary low-level features and an edge map of the fashion apparel and (b) a second level encoder generating the target style embedding from the concatenated feature embedding;
generate a target color embedding from the RGB components of the target color via multi-layer Fully Connected (FC) layers of the FAR-GAN; and
generate via a decoder of the FAR-GAN, a synthetic target image comprising the color variant of the fashion apparel, wherein the decoder dictates the target style embedding by the target color embedding using an Adaptive Instance Normalization (AdaIN) and utilizes un-correlated Gaussian noise added to each of the plurality of decoder blocks to impose variability in a color and a style of the fashion apparel in the synthetic target image enabling the synthetic target image to appear realistic when displayed on a display screen.

### Claim 7 (depends on 6)

The system of claim 6, wherein the FAR-GAN via a discriminator connected to the decoder is configured to determine whether the synthetic target image is a realistic or a non-realistic image by comparing the synthetic target image with the input image, wherein the FAR-GAN is trained on an adversarial loss to continuously enhance the synthetic target image.

### Claim 8 (depends on 6)

The system of claim 6, wherein the FAR-GAN via is configured to utilize of a Color Consistency Loss (CCL) to explicitly control color information and penalize the FAR-GAN until a color of the synthetic target image is similar to the target color, wherein the CCL utilizes a cost function L2 which is a function of the target color and a dominant color in the synthetic target image.

### Claim 9 (depends on 6)

The system of claim 6, wherein the FAR-GAN is configured to generate the target style embedding by:
extracting the plurality of primary low-level features of the fashion apparel via a first encoder of the first level encoder to generate an image feature embedding;
generating the edge map of the fashion apparel via a pre-trained Cycle GAN;
generating, via a second encoder of the first level encoder, an edge feature embedding from the edge map;
concatenating the image feature embedding and the edge feature embedding to generate the concatenated feature embedding; and
generating the target style embedding of the fashion apparel by processing the concatenated feature embedding via the second level encoder.

### Claim 10 (depends on 6)

The system of claim 6, wherein if the input image comprises a cluttered background, the one or more hardware processors are configured to apply segmentation process on the input image to detect the fashion apparel from the input image prior to processing the input image by the first level encoder.

### Claim 11

One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:
receiving, (a) an input image comprising a fashion apparel, and (b) a target color for generating a color variant of the fashion apparel, wherein the target color is represented in terms of Red Blue Green (RGB) components;
generating, via a two-step encoder of a Fashion Apparel Regeneration-Generative Adversarial Network (FAR-GAN) executed by the one or more hardware processors, a target style embedding of the fashion apparel by processing the input image, the two-step encoder comprising (a) a first level encoder generating a concatenated feature embedding representing a plurality of primary low-level features and an edge map of the fashion apparel and (b) a second level encoder generating the target style embedding from the concatenated feature embedding;
generating a target color embedding from the RGB components of the target color via multi-layer Fully Connected (FC) layers of the FAR-GAN; and
generating, via a decoder of the FAR-GAN executed by the one or more hardware processors, a synthetic target image comprising the color variant of the fashion apparel, wherein the decoder dictates the target style embedding by the target color embedding using an Adaptive Instance Normalization (AdaIN) and utilizes un-correlated Gaussian noise added to each of the plurality of decoder blocks to impose variability in a color and a style of the fashion apparel in the synthetic target image enabling the synthetic target image to appear realistic when displayed on a display screen.

### Claim 12 (depends on 11)

The one or more non-transitory machine-readable information storage mediums of claim 11, wherein the one or more instructions which when executed by the one or more hardware processors further cause use of a Color Consistency Loss (CCL) to explicitly control color information and penalize the FAR-GAN until a color of the synthetic target image is similar to the target color, wherein the CCL utilizes a cost function L2 which is a function of the target color and a dominant color in the synthetic target image.

### Claim 13 (depends on 11)

The one or more non-transitory machine-readable information storage mediums of claim 11, wherein the one or more instructions which when executed by the one or more hardware processors further cause:
extracting the plurality of primary low-level features of the fashion apparel via a first encoder of the first level encoder to generate an image feature embedding;
generating the edge map of the fashion apparel via a pre-trained Cycle GAN;
generating, via a second encoder of the first level encoder, an edge feature embedding from the edge map;
concatenating the image feature embedding and the edge feature embedding to generate the concatenated feature embedding; and
generating the target style embedding of the fashion apparel by processing the concatenated feature embedding via the second level encoder.

### Claim 14 (depends on 11)

The one or more non-transitory machine-readable information storage mediums of claim 11, wherein if the input image comprises a cluttered background, a segmentation process is applied on the input image to detect the fashion apparel from the input image prior to processing the input image by the first level encoder.
