Generative Adversarial Network Training for Super-Resolution Image Detail

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current super-resolution image reconstruction methods often prioritize small reconstruction distortion over perceptual distortion, resulting in images that lack detail and appear unreal.

Innovation Solution

A training method for generative adversarial networks that includes a generative network and a discriminative network, where the generative network converts a first-resolution image into a second-resolution image with controlled artificial details, and the discriminative network evaluates the output to adjust the generative network's parameters, combining reconstruction and perceptual distortion evaluation criteria.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If current super-resolution reconstruction methods are used to prioritize small reconstruction distortion, then the reconstruction accuracy is improved, but the perceptual quality and detail richness deteriorate

Engineering Contradiction:
Improvereconstruction accuracyVSAvoidperceptual quality
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

The patent changes the optimization parameters from purely reconstruction-based metrics (PSNR, SSIM) to a combined metric that includes perceptual distortion. The loss function is modified to balance reconstruction accuracy with perceptual quality, allowing the model to generate images that are both accurate and visually realistic with rich details.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a feedback mechanism where the perceptual distortion evaluation is continuously fed back into the training process. The discriminative network provides feedback on perceptual quality, enabling the generative network to iteratively improve both reconstruction accuracy and perceptual quality through adversarial training.

Inventive Principle:
Principle #23Feedback

2Productivity

If traditional super-resolution methods are used, then the processing speed is maintained, but the image detail information and realism deteriorate

Engineering Contradiction:
Improveprocessing speedVSAvoidimage realism
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent replaces traditional mechanical image processing methods with a neural network-based generative model. The generative adversarial network uses learned transformations and adversarial optimization to produce realistic high-resolution images, substituting conventional algorithms with an intelligent system that achieves both speed and quality.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the fundamental parameters of image processing by using deep learning models with millions of parameters to capture complex relationships between low-resolution and high-resolution images. This enables the system to generate realistic details while maintaining processing efficiency through optimized network architectures.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3859655B1Training method, image processing method, device and storage medium for generative adversarial network
Publication Date: 2025.04.23 BOE TECHNOLOGY GROUP CO LTD
  • EP3859655B1 patent drawingFigure 1~2
  • EP3859655B1 patent drawingFigure 3
  • EP3859655B1 patent drawing

AI summary

The present disclosure provides a training method for generative adversarial network including a generative network training procedure, which includes: extracting a first-resolution sample image from a second-resolution sample image; separately providing a first input image and a second input image for a generative network to generate a first output image and a second output image respectively, the first input image including a first-resolution sample image and a first noise image corresponding to a noise sample with a first amplitude, the second input image including the first-resolution sample image and a second noise image corresponding to a noise sample with a second amplitude, a first amplitude being greater than 0, and a second amplitude being equal to 0; separately providing the first output image and a second-resolution sample image for a discriminative network to output a first discrimination result and a second discrimination result; and adjusting parameters of the generative network to reduce a loss function. The present disclosure further provides an image processing method using the generative adversarial network, a computer device, and a computer-readable storage medium. The present disclosure can obtain images which can meet required needs.