Scale-Independent Blur Kernel Estimation for Blind Super-Resolution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional deep learning-based super-resolution models are designed for pre-defined blur kernel-based bicubic degradation, which deteriorates performance due to the complexity and unknown nature of real-world blur kernels, and existing blind SR models fail to generate realistic low-resolution images.

Innovation Solution

A fully unsupervised Generative Adversarial Network (GAN)-based deep neural network is used to estimate a scale-independent blur kernel, generating realistic low-resolution images by optimizing a model with pixel-wise loss functions and estimating a range of Gaussian blur standard deviation, allowing for the creation of training pairs for deep convolutional super-resolution networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If pre-defined blur kernel-based bicubic degradation is used for super-resolution training, then the training process is simple and straightforward, but the performance deteriorates due to the complexity and unknown nature of real-world blur kernels

Engineering Contradiction:
Improvetraining process simplicityVSAvoidsuper-resolution performance
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The model performs self-service by automatically estimating its own blur kernel from the input low-resolution image without requiring external ground truth or manual specification. The blur kernel estimation module extracts degradation characteristics directly from the degraded image, enabling the system to adapt to real-world blur conditions while maintaining training simplicity

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system dynamically changes the blur kernel parameters based on the specific input image rather than using fixed pre-defined kernels. The blur kernel estimation module computes image-specific kernel parameters (such as standard deviation for Gaussian kernels) that match the actual degradation, allowing the super-resolution model to adapt to varying blur conditions

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If blind super-resolution models are used to handle unknown blur kernels, then the adaptability to real-world images improves, but the ability to generate realistic low-resolution images deteriorates

Engineering Contradiction:
Improveadaptability to unknown blurVSAvoidrealism of generated low-resolution images
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The system implements feedback through the blur kernel estimation module that continuously analyzes the generated low-resolution images and adjusts the kernel parameters accordingly. The estimated blur kernel is fed back into the degradation process to ensure consistency between the generated LR images and the actual blur characteristics, improving realism while maintaining adaptability

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The blur kernel estimation is performed as a preliminary action before the super-resolution reconstruction. By estimating the blur kernel from the input LR image first, the system prepares accurate degradation parameters that guide the subsequent HR image generation, ensuring both adaptability to unknown blur and realism in the output

Inventive Principle:
Principle #10Preliminary action

3Ease of manufacture

If scale-dependent blur kernels are used for training, then the training data generation is straightforward, but the model performance deteriorates on images with different scale factors

Engineering Contradiction:
Improvetraining data generationVSAvoidperformance across different scales
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The blur kernel estimation module serves multiple functions: it estimates blur parameters for training data generation, adapts to different scale factors, and handles various blur types (Gaussian, motion, etc.). This universal approach allows the system to generate realistic training data across different scales without requiring separate training procedures for each scale factor

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4272152B1Unsupervised learning-based scale-independent blur kernel estimation for super-resolution
Publication Date: 2026.03.11 SAMSUNG ELECTRONICS CO LTD
  • EP4272152B1 patent drawingFigure 1
  • EP4272152B1 patent drawingFigure 2A
  • EP4272152B1 patent drawingFigure 2B

AI summary

One embodiment provides a method generating a first image crop and a second image crop randomly extracted from a low-quality image and a high-quality image, respectively. The method further comprises comparing the first image crop and the second image crop using a plurality of loss functions including pixel-wise loss to calculate losses, and optimizing a model trained to estimate a realistic scale-independent blur kernel of a low-resolution (LR) blurred image by minimizing the losses.