Feature-Preserving Noise Removal via Multi-Copy Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional noise removal techniques often fail to effectively remove noise from data without introducing artifacts or reducing information content, especially when noise distributions are arbitrary or unknown, and they struggle with high noise levels and multiplicative noises.
Innovation Solution
A noise removal system using a multi-copy multi-layer perceptron (MCMLP) technique that transforms data into input vectors and processes them through multiple layers of perceptrons, allowing for noise estimation and reconstruction without prior assumptions about noise statistics, capable of handling arbitrary noise distributions and extreme noise levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional noise removal techniques are applied, then noise is reduced, but artifacts are introduced and information content is lost
Solution Approach 1:
The patent segments the noisy signal into multiple independent copies or realizations. By processing these segmented copies through the neural network separately and then combining their outputs, the method preserves more information while removing noise, avoiding the information loss that occurs in conventional single-pass filtering approaches.
Solution Approach 2:
The patent introduces a neural network as an intermediary between the noisy input and the final denoised output. This intermediary learns the complex mapping from noisy to clean signals during training, enabling it to remove noise while preserving information content better than direct conventional filtering methods.
2Object-affected harmful factors
If conventional noise removal techniques are applied, then noise is reduced, but artifacts are introduced
Solution Approach 1:
By segmenting the input into multiple copies and processing them independently through the neural network, the method avoids the artifact generation that occurs in conventional techniques. The segmentation allows each copy to be processed with the learned denoising model, and their combination produces artifact-free results.
Solution Approach 2:
The trained neural network acts as an intermediary that has learned to distinguish between noise and actual signal features during training. This intermediary produces artifact-free outputs by applying the learned denoising transformation rather than using fixed conventional filters that generate artifacts.
3Device complexity
If conventional noise removal techniques are applied, then processing is simple, but they fail with arbitrary noise distributions and high noise levels
Solution Approach 1:
The patent applies preliminary action by training the neural network offline on clean training data before actual noise removal is needed. This pre-training phase captures the noise characteristics and denoising transformations, making the actual noise removal process simple and reliable even for arbitrary noise distributions that were not explicitly programmed.
Solution Approach 2:
The patent changes the parameters of the processing system by using a trained neural network with learned weights and biases instead of fixed conventional filter parameters. This allows the system to adapt to arbitrary noise distributions and high noise levels while keeping the actual processing simple, as the complex adaptation has already been performed during training.
Data Source
AI summary
A noise removal method includes receiving in a de-noising system multiple noisy copies of a target image; providing the noisy copies as input data to the de-noising system in a training phase, and determine a noise distribution for the plurality of noisy copies. The method further includes, for n stages of the de-noising system, transforming the input data into an output; reconstructing the output; and providing the reconstructed output as input data. The method may further include receiving a new image and generating a de-noised image based on the determined noise distribution.