Texture-Aware Blind Denoising with Self-Residual Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional self-supervised blind denoising methods fail to effectively eliminate spatially correlated noise in real-world images without degrading texture details, as they rely on Pixel-shuffle Downsampling (PD) processes that destroy high-frequency details.

Innovation Solution

A blind denoising device and method that uses self-residual learning to eliminate noise without PD, employing a receiving unit, a Pixel-shuffle Downsampling (PD) unit, a predicted image producing unit, and a learning unit with self-supervised losses to restore high-frequency characteristics and eliminate spatially correlated noise.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If Pixel-shuffle Downsampling (PD) is used to eliminate spatially correlated noise, then noise elimination performance is improved, but texture details are degraded

Engineering Contradiction:
Improvenoise elimination performanceVSAvoidtexture detail quality
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The image processing is segmented into two distinct phases: training phase where PD is applied to generate diverse downsampled images for robust noise pattern learning, and testing phase where PD is completely removed to preserve original texture details. This segmentation allows each phase to optimize for its specific goal without compromising the other.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model is preliminarily trained using downsampled images with PD during the training phase, enabling the network to learn noise patterns and restoration capabilities in advance. This preliminary action equips the model with noise elimination skills that can be applied later without actually applying PD during the final denoising operation.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If PD-based inference is used in the testing phase, then noise elimination is enhanced, but high-frequency details are destroyed

Engineering Contradiction:
Improvenoise elimination capabilityVSAvoidhigh-frequency detail loss
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The model serves itself by internally learning noise patterns during training without requiring external PD operations during testing. The network develops self-sufficient noise elimination capabilities through exposure to downsampled training data, enabling it to function effectively in the testing phase without actually applying PD, thus preserving high-frequency information.

Inventive Principle:
Principle #25Self-service

3Manufacturing precision

If a model without PD process is used, then texture details are preserved, but noise elimination performance decreases

Engineering Contradiction:
Improvetexture detail preservationVSAvoidnoise elimination performance
Core Design Contradiction:
Manufacturing precisionVSReliability

Solution Approach 1:

During training, the model processes copies of the original image that have been downsampled via PD. These copied and transformed versions help the network learn noise patterns without permanently altering the original image quality. The model learns from these copies while preserving the ability to restore the original image details in the testing phase.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4386660B1Device and method for texture-aware self-supervised blind denoising using self-residual learning
Publication Date: 2025.08.06 KOREA UNIV RES & BUSINESS FOUND
  • EP4386660B1 patent drawingFigure 1~3
  • EP4386660B1 patent drawingFigure 4~5
  • EP4386660B1 patent drawingFigure 6

AI summary

The present invention relates to a blind denoising device including: a receiving unit for receiving an original noisy image; a Pixel-shuffle Downsampling (PD) unit for performing PD for the original noisy image to produce at least one or more downsampled images; a predicted image producing unit for eliminating the noise from the original noisy image and the downsampled images to produce at least one or more predicted images restored to the shape of the original noisy image; and a learning unit for performing optimized learning for the predicted image producing unit, based on at least one or more self-supervised losses of the predicted images and the original noisy image.