Super Resolution Image Processing Using Adaptive Supervisory Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution technologies using machine learning struggle to accurately infer high-frequency components in images due to limited similarity between training data and inference target images, leading to reduced inference accuracy, especially when conditions like weather or illuminance differ.
Innovation Solution
An image processing apparatus and method that selects pairs of supervisory data from a first image group to generate a learning model for a second image group with fewer high-frequency components, using frames captured at identical times to improve inference accuracy by creating a learning model tailored to the target image's conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large number of various images are used as supervisory data to ensure equal accuracy for all inference target moving images, then the adaptability of the learning model improves, but the inference accuracy for specific target images decreases due to low similarity between training data and target images
Solution Approach 1:
The patent applies local quality by selecting supervisory data that is specifically similar to the inference target image in terms of image patterns, rather than using uniformly diverse data. This means different parts of the data selection process focus on different qualities: diversity for adaptability, but similarity for accuracy. The system dynamically adjusts the characteristics of supervisory data based on the specific inference target, making the data quality locally optimized for each inference task.
Solution Approach 2:
The patent implements dynamics by making the supervisory data selection adaptive and dynamic rather than static. The system determines supervisory data based on the specific inference target image, allowing the characteristics of training data to change dynamically according to the target. This dynamic adaptation enables the system to optimize both adaptability and accuracy by adjusting data selection criteria based on real-time requirements.
2Quantity of substance
If images captured at different locations and times are used as supervisory data, then the quantity of training data increases, but the inference accuracy decreases due to differences in objects captured
Solution Approach 1:
The patent applies parameter changes by selectively adjusting the parameters of supervisory data based on similarity to the inference target. Rather than uniformly increasing data quantity, the system changes the selection parameters to prioritize images with similar patterns, objects, and characteristics to the target image. This selective parameter adjustment ensures that quantity increase does not compromise accuracy.
3Measurement precision
If learning is performed using only images from a specific section with high similarity to the inference target, then the inference accuracy for that section improves, but the accuracy decreases for other sections due to limited supervisory data variety
Solution Approach 1:
The patent implements dynamics by making the supervisory data selection adaptive to the inference target rather than fixed to a specific section. The system dynamically determines which images to use as supervisory data based on similarity assessment, allowing it to optimize for specific sections when needed while maintaining generalization capability through flexible data selection. This dynamic approach resolves the contradiction between specialized accuracy and general adaptability.
Data Source
AI summary
An image processing apparatus configured to achieve high definition of an image of a second image group by using a first image group, the second image group having less high-frequency components in a corresponding frame than the first image group, the image processing apparatus comprises a selection unit configured to select, based on a high-definition target image selected from the second image group, a pair of supervisory data to be used for learning, a learning model generation unit configured to generate a learning model by using the pair of supervisory data, an inference unit configured to infer high-frequency components of the high-definition target image by using the learning model generated, and an image generation unit configured to generate a high-definition image based on the high-definition target image and the high-frequency components.