Convolutional Neural Network Training for High-Frequency Denoising

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image denoising methods using convolutional neural networks often lose high frequency information, resulting in blurred images that lack small details and fail to enhance the signal-to-noise ratio (SNR).

Innovation Solution

Applying a high pass filter to compare denoised images with original noisy images, training the neural network to recognize and reinject high frequency patterns missing in denoised images, using a high frequency loss function to minimize the loss of convolutional neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If image denoising is performed using conventional convolutional neural networks, then noise is reduced, but high frequency information is lost resulting in blurred images

Engineering Contradiction:
ImprovenoiseVSAvoidhigh frequency information
Core Design Contradiction:
Object-affected harmful factorsVSLoss of information

Solution Approach 1:

The patent segments the loss function into multiple components: a conventional loss term for noise reduction and a high frequency loss term for preserving fine details. This segmentation allows the network to independently optimize for both noise removal and high frequency preservation without one objective dominating the other, thereby resolving the contradiction between reducing noise and preserving high frequency information.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a high frequency loss function that operates in the frequency domain, transforming how the network evaluates denoising performance. By changing the parameter space from purely spatial domain error metrics to frequency-domain based metrics, the network can maintain high frequency patterns while removing noise, thus resolving the information loss problem.

Inventive Principle:
Principle #35Parameter changes

2Use of energy by moving object

If exposure time or gain is reduced to minimize photo bleaching and photo toxicity, then energy input is minimized, but image quality deteriorates with increased noise

Engineering Contradiction:
Improveenergy inputVSAvoidnoise
Core Design Contradiction:
Use of energy by moving objectVSObject-affected harmful factors

Solution Approach 1:

The patent converts the harmful noise present in low-exposure images into a useful training signal. By using the noisy low-exposure images as input and training the network to produce high-quality output, the noise that would normally be considered a drawback becomes part of the learning process, enabling the network to learn effective denoising strategies that preserve detail while working with low-energy acquisition parameters.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The convolutional neural network acts as an intermediary between the noisy low-exposure image and the desired high-quality output. Rather than directly capturing high-quality images requiring high energy input, the network mediates the transformation from noisy low-exposure images to clean high-frequency preserving images, enabling energy-efficient imaging without sacrificing quality.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Object-affected harmful factors

If conventional denoising methods are applied, then noise is removed, but structural similarity and detail preservation are reduced

Engineering Contradiction:
ImprovenoiseVSAvoidstructural similarity
Core Design Contradiction:
Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The patent moves the evaluation from the spatial domain to the frequency domain by introducing a high frequency loss function. This dimensional change allows the network to preserve high frequency patterns that are critical for structural similarity and detail preservation, while still achieving effective noise removal in the spatial domain.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The high frequency loss function provides feedback to the network during training, specifically highlighting areas where high frequency information has been lost. This feedback mechanism guides the network to adjust its denoising operations to preserve edges, textures, and fine details, thereby maintaining structural similarity while removing noise.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3996035B1Methods and systems for training convolutional neural networks
Publication Date: 2025.07.23 LEICA MICROSYSTEMS CMS GMBH
  • EP3996035B1 patent drawingFigure 1
  • EP3996035B1 patent drawingFigure 2~3
  • EP3996035B1 patent drawingFigure 4~5

AI summary

A computer implemented method for training a convolutional neural network is described. The method comprises receiving a captured image; generating a denoised image by applying the convolutional neural network to the captured image; and training the convolutional neural network based on a high frequency loss function, the captured image and the denoised image.