Image Super-Resolution Using Local Similarity-Based Region Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution imaging techniques using machine learning struggle with low inference accuracy due to insufficient similarity between teacher data and inference target images, particularly when imaging conditions or locations vary, leading to reduced performance in generating high-frequency components.
Innovation Solution
An image processing apparatus that selects teacher data based on similarity with a current image, divides the image into local regions, generates a learning model for each region, and infers high-frequency components to enhance image definition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a learning model is generated using a wide variety of images as teacher data, then the model can handle diverse image types, but the inference accuracy for specific images with low similarity in the teacher data is reduced
Solution Approach 1:
The patent divides the image into multiple local regions and generates separate learning models for each region based on local characteristics. This segmentation allows the system to maintain versatility across different image types while achieving high inference accuracy for each specific region by using locally relevant teacher data.
Solution Approach 2:
The patent applies local quality by generating learning models tailored to specific local regions of the image. Each learning model is trained on teacher data that is locally similar to the corresponding image region, ensuring high inference accuracy for that specific area while the overall system maintains capability to handle diverse image types through region-specific customization.
2Measurement precision
If learning is performed using teacher data with similar imaging location and conditions, then inference accuracy for that specific location is improved, but the model's ability to generalize to other locations is reduced
Solution Approach 1:
The patent segments the imaging area into multiple local regions and creates separate learning models for each. This allows the system to train on location-specific teacher data for high accuracy in each region while maintaining the ability to handle various locations through the distributed model structure.
Solution Approach 2:
The patent achieves universality by creating a multi-functional learning system where multiple learning models can be generated for different locations and conditions. Each model is specialized for its target location but the overall system can adapt to various imaging locations and conditions through the selection and application of appropriate regional models.
3Device complexity
If a single learning model is generated for the entire image, then the processing is simple, but the inference accuracy for local regions with different characteristics is reduced
Solution Approach 1:
The patent applies segmentation by dividing the image into multiple local regions and generating separate learning models for each region. This increases processing complexity but significantly improves inference accuracy for local regions with different characteristics, as each region receives customized training data and model parameters.
Solution Approach 2:
The patent implements local quality by creating learning models with region-specific characteristics. Each learning model is trained on teacher data matching the local region's characteristics, enabling high inference accuracy for each area while the system manages complexity through structured regional organization.
Data Source
AI summary
An image processing apparatus calculates a degree of similarity with a partial region corresponding to a previous image which is a high definition target previous to a current image selected as a high definition target for each one of a plurality of partial regions obtained by dividing the current image, determines a plurality of local regions form the current image by combining a collection of one or more partial regions with the degree of similarity equal to or greater than a threshold as one local region and treating a partial region with the degree of similarity less than the threshold as a separate local region. The image processing apparatus infers high frequency components for each one of the plurality of local regions using a learning model selected based on the current image.