Generative Adversarial Network Conditional Discriminator for High-Resolution Image Synthesis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Generative adversarial networks (GANs) face challenges in generating high-resolution images, particularly when dealing with varying input data, such as medical images, where the features of generated images are averaged, and the network architecture needs to be extensively corrected, leading to mode collapse and limited variation in generated images.

Innovation Solution

A learning method for a generative model that estimates high-resolution images from low-resolution inputs by using a generative adversarial network with a generator that only inputs the low-resolution image and a discriminator that receives part and slice information, allowing for learning without adjusting network parameters or architecture, enabling the generation of high-resolution images across various conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If category information is input to the generator to adjust the generative model, then the model can be adjusted based on category information, but the network architecture needs to be largely corrected and the generator interface becomes complicated

Engineering Contradiction:
Improveadjustability of generative modelVSAvoidnetwork architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent introduces a conditional discriminator as an intermediary component that receives category information and uses it to guide the generation process. The discriminator outputs conditional information that is fed back to the generator, allowing the generator to adapt to different categories without modifying its architecture. This mediator approach enables category-based adjustment while keeping the generator interface simple and architecture unchanged.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If the GAN learns from various input data, then it can handle different data types, but the features of the generated image are averaged and mode collapse occurs

Engineering Contradiction:
Improvehandling of various input dataVSAvoidimage generation accuracy
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The patent implements dynamic conditional discrimination where the discriminator adapts its judgment criteria based on the input category. The conditional discriminator dynamically adjusts which features to emphasize in its evaluation based on the category information, allowing the GAN to maintain high generation accuracy for each specific category while still handling diverse input data types. This dynamic adaptation prevents feature averaging and mode collapse.

Inventive Principle:
Principle #15Dynamics

3Manufacturing precision

If the generator produces high resolution images from low resolution inputs, then detailed images can be generated, but parameter adjustment and network architecture correction are required

Engineering Contradiction:
Improveimage resolutionVSAvoidparameter adjustment burden
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent employs self-service learning where the generator automatically learns to produce high-resolution images from low-resolution inputs through the guidance of the conditional discriminator. The system performs self-adjustment during the learning process without requiring manual parameter tuning or architecture modification. The conditional discriminator provides category-specific feedback that guides the generator to automatically adapt its parameters and structures to achieve high-resolution generation for different categories.

Inventive Principle:
Principle #25Self-service

Data Source

PatentUS11948349B2Learning method, learning device, generative model, and program
Publication Date: 2024.04.02 FUJIFILM CORP
  • US11948349B2 patent drawing
  • US11948349B2 patent drawing
  • US11948349B2 patent drawing

AI summary

Provided are a learning method, a learning device, a generative model, and a program that generate an image including high resolution information without adjusting a parameter and largely correcting a network architecture even in a case in which there is a variation of the parts of an image to be input. Only a first image is input to a generator of a generative adversarial network that generates a virtual second image having a relatively high resolution by using the first image having a relatively low resolution, and a second image for learning or the virtual second image and part information of the second image for learning or the virtual second image are input to a discriminator that identifies the second image for learning and the virtual second image.