Neural Network Upscaling Synthetic Data Rendering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generating high-resolution image and video content is resource-intensive and often results in artifacts, making it difficult to achieve desired quality, especially for devices with limited capacity, and obtaining sufficient training data for neural networks is challenging.
Innovation Solution
A deep learning-based approach that uses a renderer to generate low-resolution images, which are then upscaled using a neural network with a Gaussian filter and blending components to produce high-resolution images, leveraging synthetic training data to avoid artifacts and improve rendering efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If high-resolution content is generated using traditional rendering methods, then image quality is improved, but resource consumption increases and frame rate decreases
Solution Approach 1:
The system pre-trains neural networks using synthetic training data generated from low-resolution images. This preliminary training enables the network to learn upscaling patterns beforehand, allowing real-time high-resolution generation without intensive rendering during actual content delivery, thus achieving high frame rates while maintaining quality
Solution Approach 2:
The patent replaces traditional mechanical rendering processes with a neural network-based upscaling system. Instead of rendering high-resolution images directly through computationally intensive graphical processing, the system uses trained neural networks to transform low-resolution images into high-resolution outputs, significantly reducing processing resources and increasing frame rates
2Productivity
If traditional upscaling methods are used to generate high-resolution content, then resource consumption is reduced, but image quality deteriorates due to artifacts
Solution Approach 1:
The system performs preliminary training of neural networks using synthetic training data that includes low-resolution images and their corresponding high-resolution targets. This advance preparation enables the network to learn complex upscaling transformations, allowing artifact-free high-resolution generation with minimal real-time processing resources
Solution Approach 2:
The patent replaces conventional upscaling algorithms with a neural network-based approach. The neural network learns to predict high-resolution details from low-resolution inputs through training on synthetic data, eliminating the artifacts produced by traditional interpolation methods while maintaining low resource consumption during operation
3Measurement precision
If real training data is collected for neural network training, then training accuracy is improved, but data acquisition time and complexity increase
Solution Approach 1:
The system creates synthetic training data by rendering low-resolution images using a rendering engine and generating corresponding high-resolution versions through controlled processes. These synthetic copies serve as training pairs for the neural network, eliminating the need to collect real-world high-resolution images while providing sufficient training accuracy
Solution Approach 2:
The patent replaces the mechanical process of collecting and annotating real training images with an automated synthetic data generation system. The rendering engine generates pairs of low-resolution and high-resolution images algorithmically, eliminating time-consuming data collection, cleaning, and annotation processes while providing adequate training data
Data Source
AI summary
Apparatuses, systems, and techniques are presented to train one or more neural networks. In at least one embodiment, one or more neural networks are trained based, at least in part, on two or more versions of an image, wherein each of the two or more versions of the image are to be synthetically generated independently.