Multi-Resolution Image Generation for Super-Resolution Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining high-quality training data for super-resolution models is challenging, particularly for different end-use scenarios, as conventional systems are time-intensive and do not generate images at multiple resolutions simultaneously.
Innovation Solution
Modify existing image generators, such as gaming engines, to simultaneously produce pairs of images at different resolutions, enabling the generation of large volumes of high-quality training data by pairing low-resolution and high-resolution images for training super-resolution models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional image processing systems are used to generate training data, then images can be rendered at different resolutions, but the process is time-intensive and does not generate multiple resolutions simultaneously
Solution Approach 1:
The image generator is modified to segment the rendering process into multiple simultaneous output streams, each producing images at different resolutions. This allows the system to generate low-resolution and high-resolution images in parallel from the same input frame data, dramatically improving training data generation efficiency while reducing the time required.
Solution Approach 2:
The image generator is enhanced to perform multiple functions simultaneously - it now generates images at multiple resolutions (low-resolution and high-resolution) from a single input frame. This multi-functionality allows the same rendering engine to produce diverse training data without requiring separate processing pipelines, thus improving productivity without proportionally increasing time consumption.
2Manufacturing precision
If high-quality training data is obtained through conventional methods, then model training accuracy can be improved, but computational costs and time consumption increase
Solution Approach 1:
The system performs preliminary rendering at high resolution and then generates low-resolution versions from the same high-resolution output. This preliminary action ensures that high-quality training data is obtained while avoiding the need for separate expensive high-resolution rendering processes, thereby reducing overall computational costs.
Solution Approach 2:
Instead of generating independent high-resolution images through separate rendering processes, the system creates copies of the high-resolution image at lower resolutions. This copying approach maintains the quality of the source data while significantly reducing the computational energy required compared to generating all resolutions independently.
3Productivity
If image generators are modified to output multiple resolutions simultaneously, then training data generation efficiency improves, but device complexity increases
Solution Approach 1:
The image generator is extended into another dimension of output capability by adding resolution as an additional output parameter. Instead of processing images sequentially at different resolutions, the system now operates in a multi-resolution output space, generating multiple resolution variants simultaneously from each input frame, thus improving efficiency without requiring fundamentally new processing architectures.
Data Source
AI summary
Systems are provided for generating training data from images that are obtained from image generators that are typically configured to only generate a single image per frame. The image generators are modified or otherwise controlled to generate two different images at different resolutions for each of a plurality of frames. The training data is created by pairing the low-resolution images and high-resolution images for common frames into training data set pairings. A super-resolution model is applied to the training data set pairings to create a trained super-resolution model.