Super-Resolution Neural Network Residual Feedback Loop

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing super-resolution image generation methods using standard interpolation algorithms often fail to capture detailed features of the original image, resulting in a lack of image fidelity in the super-resolution output.

Innovation Solution

The proposed method involves an image processing system that uses an upscaling convolutional network to generate a feature image, followed by a downscaling convolutional network to produce a downscaled feature image. A residual image is then determined between the downscaled and original feature images, upscaled, and used to correct the original feature image, ultimately generating a high-quality super-resolution image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Length of stationary object

If standard interpolation algorithms are used to generate super-resolution images, then the image resolution is increased, but the detailed features of the image are lost

Engineering Contradiction:
Improveimage resolutionVSAvoiddetailed features
Core Design Contradiction:
Length of stationary objectVSLoss of information

Solution Approach 1:

The patent implements a feedback mechanism by downsampling the upscaled feature image and comparing it with the original feature image to generate a residual image. This residual information is then fed back through an upsampling network and added to the original upscaled features, creating a closed-loop system that continuously refines the super-resolution output to preserve detailed features.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces an intermediary residual image that captures the difference between the downscaled and original feature images. This residual image serves as a mediator that carries detailed feature information, which is then upscaled and integrated with the main upscaled features to enhance the final super-resolution image quality.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If deep Laplacian Pyramid Super-Resolution Network is used, then image super-resolution is achieved, but the computational complexity increases

Engineering Contradiction:
Improveimage qualityVSAvoidnetwork structure
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the super-resolution task into distinct functional modules: an upsampling network for initial resolution enhancement, a downsampling network for feature extraction, a residual network for detail enhancement, and a synthesis network for final image generation. This modular segmentation reduces overall computational complexity while maintaining high image quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by focusing computational resources on specific aspects of image enhancement. Instead of processing the entire image uniformly through multiple complex stages, the residual network selectively processes only the difference information (residuals) that contain the most important detailed features, reducing unnecessary computations.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3776451B1Image processing apparatus, image processing method thereof, image processing system, and training method thereof
Publication Date: 2025.02.12 BOE TECHNOLOGY GROUP CO LTD
  • EP3776451B1 patent drawingFigure 1
  • EP3776451B1 patent drawingFigure 2
  • EP3776451B1 patent drawingFigure 3

AI summary

The present disclosure relates to an image processing method. The image processing method may include upscaling a feature image of an input image by an upscaling convolutional network to obtain a upscaled feature image; downscaling the upscaled feature image by a downscaling convolutional network to obtain a downscaled feature image; determining a residual image between the downscaled feature image and the feature image of the input image; upscaling the residual image between the downscaled feature image and the feature image of the input image to obtain an upscaled residual image; correcting the upscaled feature image using the upscaled residual image to obtain a corrected upscaled feature image; and generating a first super-resolution image based on the input image using the corrected upscaled feature image.