Image Enhancement Model Noise Distribution Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current image enhancement technologies, such as deep learning-based neural networks, face challenges in accurately estimating noise distribution and effectively removing noise from images, especially when capturing parameters like ISO values and Bayer pattern information are not fully considered.

Innovation Solution

A processor-implemented method that estimates noise distribution using a noise model trained on pixel data, pixel position information, and capturing parameters, and then generates an enhanced image using an image enhancement model that accounts for these factors, allowing for improved noise removal and image quality enhancement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If deep learning-based neural networks are used for image enhancement, then image quality can be improved, but accurate noise distribution estimation becomes difficult when capturing parameters are not considered

Engineering Contradiction:
Improveimage qualityVSAvoidnoise distribution estimation
Core Design Contradiction:
Manufacturing precisionVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by training the neural network model in advance with capturing parameters (ISO, exposure time, Bayer pattern) integrated into the training process. This allows the model to pre-learn the relationship between capturing conditions and noise characteristics, enabling accurate noise distribution estimation during actual image enhancement without requiring real-time parameter analysis.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements parameter changes by incorporating multiple capturing parameters (ISO value, exposure time, Bayer pattern information) as additional input dimensions to the neural network. This transforms the noise estimation problem from a single-image analysis to a multi-parameter joint analysis, allowing the model to adapt noise estimation based on the specific capturing conditions that generated the input image.

Inventive Principle:
Principle #35Parameter changes

2Productivity

If noise removal is performed without considering capturing parameters, then processing speed can be maintained, but noise removal accuracy deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidnoise removal accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent uses preliminary action by pre-training the neural network with capturing parameter information during the model training phase. This allows the model to internally store knowledge about how different capturing conditions affect noise characteristics, enabling fast inference during actual processing without requiring complex real-time parameter analysis, thus maintaining both speed and accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces capturing parameters as an intermediary element that mediates between the input image and the noise estimation process. These parameters serve as additional contextual information that helps the neural network distinguish between different noise types and intensities, improving noise removal accuracy while the efficient neural network architecture maintains processing speed.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4105876A1Method and apparatus with image enhancement
Publication Date: 2022.12.21 SAMSUNG ELECTRONICS CO LTD
  • EP4105876A1 patent drawingFigure 1
  • EP4105876A1 patent drawingFigure 2
  • EP4105876A1 patent drawingFigure 3

AI summary

Methods and apparatuses with training or image enhancement are disclosed. The image enhancement method includes obtaining an input image, estimating a noise distribution of the input image by implementing a noise model based on the input image, and generating an enhanced image by implementing an image enhancement model dependent on the input image and the estimated noise distribution.