Convolutional Neural Network Training Using Statistical Noise Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for image denoising struggle to effectively remove noise while preserving image features, particularly in low exposure time or gain images contaminated with additive noise.

Innovation Solution

A computer-implemented method for training a convolutional neural network (CNN) that involves generating a statistical noise model based on a captured image and using this model to train the CNN, allowing it to learn noise reduction without relying on traditional training images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use 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 due to increased noise

Engineering Contradiction:
Improveenergy inputVSAvoidimage quality
Core Design Contradiction:
Use of energy by moving objectVSReliability

Solution Approach 1:

The patent converts the harmful noise present in low-exposure images into useful training data. By treating noisy images as training samples for a convolutional neural network, the previously harmful noise becomes the basis for teaching the system to recognize and remove noise patterns, thereby transforming the disadvantage of low signal-to-noise ratio into a benefit for training robust denoising models

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

Solution Approach 2:

The patent creates synthetic training pairs by applying the same noise model to both original and denoised images. Multiple copies of training data are generated through data augmentation techniques including random patches, swapping pixel values with neighbor pixels, and creating probability distributions based on statistical noise models, allowing the network to learn from diverse variations without requiring additional physical images

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If traditional training images are used to train the CNN, then training data availability is limited, but the method cannot adapt to real acquired images with different noise characteristics

Engineering Contradiction:
Improveadaptability to different imagesVSAvoidtraining data
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The system performs self-training by using its own output as training data. The convolutional neural network generates denoised images, which are then used as ground truth to train subsequent iterations of the network. This self-service mechanism allows continuous adaptation to real acquired images without requiring external training datasets, enabling the system to improve its performance autonomously on the specific images it processes

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent performs preliminary statistical analysis of the input image to characterize noise properties before training begins. By pre-computing variance or standard deviation values and generating a statistical noise model from the captured image, the system prepares adaptive training data that reflects the actual noise characteristics of the real images, enabling the network to be pre-adapted to the specific imaging conditions before full training commences

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3996034B1Methods and systems for training convolutional neural networks
Publication Date: 2025.04.16 LEICA MICROSYSTEMS CMS GMBH
  • EP3996034B1 patent drawingFigure 1
  • EP3996034B1 patent drawingFigure 2
  • EP3996034B1 patent drawingFigure 3

AI summary

A computer implemented method for training a convolutional neural network is described. The method comprising receiving a captured image; generating a statistical noise model based on the captured image; and training a convolutional neural network based on the captured image and the statistical model.