Endoscope Image Super-Resolution With Nyquist Frequency Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for image super-resolution face challenges in generating accurate high-resolution images due to the inclusion of false patterns caused by high-frequency components not present in the training images, leading to incorrect inferences, particularly in medical imaging.
Innovation Solution
The method involves reducing high-frequency components in the candidate correct answer images beyond the Nyquist frequency of the training images, either through image processing or using an optical low-pass filter, to generate a correct answer image suitable for learning, thereby improving the resolving power of the machine learning model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution candidate correct answer images are used for machine learning, then the resolving power of the machine learning model is improved, but false patterns are introduced due to high-frequency components not present in training images
Solution Approach 1:
The patent extracts and removes high-frequency components from the candidate correct answer image that exceed the Nyquist frequency of the training image. This is achieved through frequency domain filtering, where only frequency components within the safe range (below Nyquist frequency) are retained, while excessive high-frequency components are discarded to prevent false pattern generation during machine learning.
Solution Approach 2:
The patent changes the frequency parameter by adjusting the cutoff frequency to match the Nyquist frequency of the training image. This parameter adjustment ensures that the correct answer image used for training contains only frequency components that can be properly represented given the training image's resolution, thereby maintaining consistency and preventing inference errors.
2Manufacturing precision
If components in frequency bands higher than Nyquist frequency are retained, then image detail is preserved, but false inferences are generated by the machine learning model
Solution Approach 1:
The patent converts the potentially harmful high-frequency components into a beneficial filtering mechanism. By intentionally removing excessive high-frequency components that would cause false patterns, the method creates a clean training dataset that prevents the machine learning model from learning false patterns, while still preserving necessary image details within the valid frequency range.
3Measurement precision
If the correct answer image includes all frequency components from high-resolution images, then resolution is maintained, but the machine learning model learns false patterns not present in training data
Solution Approach 1:
The patent performs preliminary frequency filtering on the candidate correct answer image before using it for machine learning training. This preliminary action removes high-frequency components that would otherwise be incorrectly learned by the model, ensuring that only valid frequency information within the training image's capabilities is used for training.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces false inferences and enhances the resolving power of the machine learning model, generating images with improved accuracy and contrast by ensuring the correct answer image only includes relevant frequency components, thus improving the inference performance.
Implementation Method 1
the correct answer image that is picked up by an image pickup apparatus mounting an optical low-pass filter, the optical low-pass filter being configured to reduce components in at least part of the frequency band higher than the Nyquist frequency of the training image
Data Source
AI summary
An image generation apparatus for machine learning performs reduction processing on components in at least part of a frequency band higher than a Nyquist frequency of a training image, for a candidate correct answer image with a resolution higher than the training image, and generate a correct answer image. The correct answer image and the training image are used as a pair for machine learning to improve a resolving power of an input image.