Multi-Frame Image Noise Reduction via Residual Map

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing noise reduction methods in image processing, particularly in multi-frame pipelines, often result in blurred feature edges and require large neural networks to process extensive data, leading to inefficiencies.

Innovation Solution

A method that acquires multiple frames, determines noise information, combines them into a single image, generates a residual noise map, and uses this map for noise reduction, allowing a neural network to process only the combined image for enhanced denoising performance without the need for extensive data processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If multiple frames are input into the AI algorithm for noise reduction, then noise reduction performance is improved, but data processing requirements and neural network size increase

Engineering Contradiction:
Improvenoise reduction performanceVSAvoidneural network size
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the noise reduction task into two distinct stages: (1) a fusion network that combines multiple frames into a single fused image, and (2) a denoising network that processes only the fused image. This segmentation allows the denoising network to operate on reduced data (single frame instead of multiple frames), significantly decreasing computational complexity and neural network size while maintaining effective noise reduction through the combined information from all input frames.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a fused image as an intermediary between the multi-frame input and the denoising process. This intermediary contains the essential information from all input frames in a consolidated form, allowing the denoising network to achieve effective noise reduction without directly processing all original frames, thus reducing data processing requirements and neural network complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If multiple frames are processed by the AI algorithm, then noise reduction performance is improved, but processing speed decreases

Engineering Contradiction:
Improvenoise reduction performanceVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent divides the processing into two efficient stages: frame fusion followed by single-frame denoising. This segmentation enables the system to leverage information from multiple frames for improved noise reduction while maintaining fast processing speed by ensuring the computationally intensive denoising operation is performed on a single fused image rather than multiple individual frames.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs frame fusion as a preliminary action before denoising. By consolidating information from multiple frames into a single fused image first, the system prepares the data in an optimized format that reduces the computational burden of the subsequent denoising operation, thereby improving overall processing speed while maintaining noise reduction effectiveness.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If multiple frames are combined and processed, then noise reduction performance is improved, but computational complexity increases

Engineering Contradiction:
Improvenoise reduction performanceVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the computational task into a fusion stage and a denoising stage, with each stage optimized for its specific function. The fusion network efficiently combines multiple frames into a single representation, and the denoising network processes only this fused result. This segmentation dramatically reduces computational complexity compared to applying denoising algorithms directly to all input frames, while still achieving improved noise reduction performance through the combined information.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts the essential information from multiple frames by fusing them into a single representative image. This extraction process removes redundant data while preserving the critical signal and noise characteristics needed for effective denoising, thereby reducing computational complexity without sacrificing noise reduction performance.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4358016A1Method for image processing
Publication Date: 2024.04.24 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP4358016A1 patent drawingFigure 1
  • EP4358016A1 patent drawingFigure 2
  • EP4358016A1 patent drawingFigure 3A~3B

AI summary

Method for image processing, in particular noise reduction, with acquiring multiple frames preferably by an image sensor; determining noise information of each frame; combining of the multiple frames in a combined image; determining a residual noise map for the combined image from the noise information of each frame; and reducing noise of the combined image on the basis of the residual noise map to generate the final image estimate.