Super-Resolution Training Data from Correlated Multi-Resolution Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining high-quality training data for super-resolution processing is challenging, particularly for different end-use scenarios, as existing systems are inefficient and time-intensive in generating image pairs at varying resolutions.
Innovation Solution
Modify existing image generators, such as gaming engines, to simultaneously or asynchronously generate pairs of images at different resolutions, creating large volumes of high-quality training data by correlating low-resolution and high-resolution images for super-resolution models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional image processing is used to generate training data, then image rendering at different resolutions is possible, but the process is time-intensive and inefficient
Solution Approach 1:
The system pre-generates high-resolution reference images and their corresponding downsampled low-resolution images before training begins. By preparing the training data in advance through predetermined downsampling operations, the system eliminates the need for time-consuming real-time generation during model training, thus improving productivity and reducing time loss.
Solution Approach 2:
The system creates training data by copying and transforming existing high-resolution images into low-resolution versions through controlled downsampling. This copying approach allows efficient generation of paired training data without requiring complex real-time rendering processes, thereby increasing data generation efficiency and reducing time requirements.
2Reliability
If high-quality training data is obtained for different end-use scenarios, then super-resolution model performance improves, but the process becomes more complex and time-consuming
Solution Approach 1:
The system employs a universal downsampling module that can process various image types and resolutions to generate training data for multiple end-use scenarios. This multi-functional approach allows the same processing pipeline to serve different training needs, improving model performance across diverse applications without proportionally increasing system complexity.
Solution Approach 2:
The system adjusts parameters such as downsampling ratios, image resolutions, and processing conditions to generate high-quality training data tailored to different end-use scenarios. By systematically varying these parameters rather than using complex adaptive systems, the method achieves improved model performance while maintaining manageable complexity levels.
3Quantity of substance
If existing image generators are modified to generate image pairs at different resolutions, then large volumes of high-quality training data can be created, but the image generator complexity increases
Solution Approach 1:
The system extracts the image generation and downsampling functions from a single complex image generator and separates them into independent modules. By taking out the downsampling operation as a distinct processing stage, the system can generate large volumes of training data through simple, modular operations rather than requiring complex integrated generation, thus increasing data volume without proportionally increasing generator complexity.
Data Source
AI summary
Systems and techniques are provided for obtaining and using training data for training a super-resolution model that transforms images from a first resolution to a second resolution. Initially, a target product is identified that generates target images at a first resolution with a first image generator. Style attributes of the target images are identified. With the style attributes, a training source product is also identified that is used to generate output images at the first resolution. Then, a second image generator is modified to generate output images for the training source product at both the first resolution and correlated output images at the second resolution. These images are used as training data for training the super-resolution model. Then, the trained super-resolution model is used to transform images for the target product from the first resolution to the second resolution.