Burst Image Restoration via Anchor-Based Feature Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image restoration methods face challenges in efficiently restoring high-quality images from burst images, particularly in environments with low illuminance or when individual images have deteriorated quality, due to the need for center line alignment which increases calculation load.

Innovation Solution

An image restoration apparatus that selects an anchor image from a burst image set and uses its anchor information to execute a neural network model, thereby generating a restored image without the need for center line alignment, reducing calculation load and enhancing image quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If center line alignment is performed to restore burst images, then image quality can be improved, but calculation load increases

Engineering Contradiction:
Improveimage qualityVSAvoidcalculation load
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent extracts and removes the center line alignment step from the traditional image restoration pipeline. By using anchor-based feature extraction and attention mechanisms, the method directly processes burst images without requiring precise geometric alignment, thereby reducing computational complexity while maintaining restoration quality

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the mechanical/geometric alignment process (center line alignment) with a neural network-based feature extraction and attention mechanism system. This substitution allows the system to handle image restoration through learned features rather than rigid geometric constraints, reducing calculation load

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

2Manufacturing precision

If neural network is trained for specialized image restoration, then restoration accuracy improves, but generalization ability to unseen patterns decreases

Engineering Contradiction:
Improverestoration accuracyVSAvoidgeneralization ability
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent designs a neural network architecture that performs multiple functions: feature extraction, attention mechanism, and image restoration in a unified model. The anchor-based approach allows the network to handle various image quality issues (noise, blur, low light) simultaneously, enhancing both specialization and generalization capabilities

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

Solution Approach 2:

The patent employs dynamic anchor adjustment and attention weight optimization to adapt the network to different input conditions. By changing parameters such as anchor positions, attention masks, and learning rates during inference, the model maintains high accuracy across diverse image patterns without requiring separate specialized models

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4040378B1Burst image-based image restoration method and apparatus
Publication Date: 2025.03.19 SAMSUNG ELECTRONICS CO LTD
  • EP4040378B1 patent drawingFigure 1
  • EP4040378B1 patent drawingFigure 2
  • EP4040378B1 patent drawingFigure 3

AI summary

The application concerns an image restoration method. The method includes determining an anchor image based on individual images of a burst image set, executing a feature extraction network based on the burst image set while using anchor information of the anchor image, and generating a restored image based on a feature map corresponding to an output of the feature extraction network. The application further concerns a computer program comprising instructions to perform said method and a non-transitory computer-readable storage medium storing said computer program. The application further concerns an apparatus comprising one or more processors configured for performing said method.