Super-Resolution Training Pairs for High-Fidelity Image Upscaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining high-quality training data for super-resolution processing is challenging, particularly for diverse end-use scenarios, which affects the performance of machine learning models in upscaling images.
Innovation Solution
Generate and utilize training data by modifying image generators to simultaneously produce pairs of low-resolution and high-resolution images, optionally with supplemental processing data, to train super-resolution models efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional image processing is used to render images at different resolutions, then content can be adapted to various display devices, but the images lack the detailed fidelity required for high-resolution displays
Solution Approach 1:
The patent uses a neural network to learn the mapping relationship between low-resolution and high-resolution images by training on paired data. Instead of using complex traditional image processing algorithms, the system copies the transformation pattern from training examples to achieve high-fidelity upscaling with simpler processing
Solution Approach 2:
The patent replaces conventional mechanical image processing techniques (such as traditional upscaling algorithms) with a machine learning-based neural network approach. This substitution enables the system to achieve superior image fidelity while maintaining computational efficiency through learned transformations rather than complex deterministic processing
2Adaptability or versatility
If super-resolution models are trained with diverse training data, then model performance across different end-use scenarios improves, but obtaining high-quality training data becomes difficult
Solution Approach 1:
The patent implements a self-service approach where the gaming engine itself generates the training data by rendering the same scene at multiple resolutions. The engine creates both low-resolution and high-resolution image pairs automatically during its normal operation, eliminating the need for external data collection efforts while achieving diverse training data across different gaming scenarios
Solution Approach 2:
The patent makes the gaming engine multi-functional by enabling it to serve both its primary purpose of rendering games and the secondary purpose of generating training data. The engine universally handles both game execution and model training data creation, allowing a single system to fulfill multiple functions without requiring separate dedicated systems for data acquisition
3Manufacturing precision
If high-resolution images are generated during runtime, then image quality improves, but game engine performance is impacted
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network model during the development phase using diverse training data. Once trained, the model can quickly transform low-resolution runtime images to high-resolution outputs without requiring the game engine to perform computationally intensive rendering operations during actual gameplay, thus maintaining performance
Solution Approach 2:
The patent segments the image processing task into two distinct phases: training phase (where the neural network learns transformations using pre-generated data) and inference phase (where the trained model quickly transforms runtime low-resolution images). This segmentation allows high-resolution output during runtime without impacting game engine performance by using a pre-computed transformation model rather than real-time complex rendering
Data Source
AI summary
Systems and methods are provided for obtaining and using training data for training a super-resolution model that transforms images from a first-low resolution to a second-high resolution. The methods include generating high-fidelity high-resolution images for the training data from underlying images at the second-high resolution and that have a relatively lower fidelity. The high-fidelity high-resolution images are paired with correlating low-resolution images and used to train the super-resolution model.