GAN Image Translation with Domain Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative Adversarial Networks (GANs) often generate fake images with undesired changes when translating images, such as altering hair color, resulting in additional artifacts like mustaches or eye color changes, making it difficult to achieve precise attribute modifications.
Innovation Solution
A method and system using a Generative Adversarial Network (GAN) with a generator and discriminator network trained on images associated with specific attributes, allowing for precise image translation by generating fake images that maintain desired attributes and avoiding unwanted changes, through the use of domain-specific training and loss functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If GANs are used to translate images with specific attribute changes, then the desired attribute modification is achieved, but additional undesired changes and artifacts are introduced
Solution Approach 1:
The patent segments the image translation task into multiple domain-specific transformations. Instead of translating from one domain to another directly, the system breaks down the translation into sequential steps through intermediate domains, where each step focuses on modifying specific attributes while preserving others. This segmentation allows precise control over which attributes change at each stage, preventing unwanted artifacts.
Solution Approach 2:
The patent introduces intermediate domains as mediators in the image translation process. When translating from domain A to domain B, the system uses domain C as an intermediary, performing A→C→B translations. These intermediate domains act as buffers that help maintain attribute integrity and reduce the introduction of undesired artifacts by distributing the transformation across multiple controlled steps.
2Adaptability or versatility
If GANs translate images between different domains, then attribute changes are achieved, but image integrity is compromised
Solution Approach 1:
The patent segments the domain translation process into attribute-specific transformations. Each domain is defined by specific attributes, and translations are performed by modifying only the target attributes while keeping other attributes constant. This segmentation ensures that image integrity is maintained by limiting changes to only what is necessary for the desired attribute modification.
Solution Approach 2:
The patent applies local quality by allowing different parts of the image to have different transformation properties. Specifically, certain attributes (like hair color or eye color) are allowed to change while other attributes (like facial structure or skin tone) are preserved. This selective transformation approach maintains overall image integrity while achieving the desired attribute changes.
3Adaptability or versatility
If GANs are trained for broad domain translation, then versatility is improved, but precision in controlling specific attribute changes deteriorates
Solution Approach 1:
The patent segments the translation task into multiple specialized sub-tasks, each handling specific attribute transformations. Instead of training a single GAN for broad domain translation, the system trains multiple GANs, each specialized in translating between specific domains with defined attribute differences. This segmentation enables precise control over which attributes change in each translation, while maintaining overall versatility through the combination of multiple specialized models.
Data Source
AI summary
Methods, systems, and media for discriminating and generating translated images are provided. In some embodiments, the method comprises: identifying a set of training images, wherein each image is associated with at least one domain from a plurality of domains; training a generator network to generate: i) a first fake image that is associated with a first domain; and ii) a second fake image that is associated with a second domain; training a discriminator network, using as inputs to the discriminator network: i) an image from the set of training images; ii) the first fake image; and iii) the second fake image; and using the generator network to generate, for an image not included in the set of training images at least one of: i) a third fake image that is associated with the first domain; and ii) a fourth fake image that is associated with the second domain.