Neural Network Image Resolution Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems struggle to convert low-resolution image data into high-resolution data without pre-prepared patch pairs, making it difficult to handle images with unexpected compositions.
Innovation Solution
A system utilizing a neural network that learns from image data pairs with different resolutions to acquire post-conversion image data, allowing for the conversion of low-resolution input data into high-resolution output data based on determined neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If patch pairs are pre-prepared for conversion, then conversion speed is improved, but adaptability to unexpected image compositions deteriorates
Solution Approach 1:
The system performs preliminary learning by generating synthetic patch pairs from original images at different resolutions during an offline training phase. This preliminary action creates a neural network model that can be quickly applied during runtime without requiring pre-prepared patches for every possible image composition, thus resolving the contradiction between conversion speed and adaptability.
Solution Approach 2:
The invention creates synthetic training data by copying and transforming original images into multiple resolutions to generate patch pairs for neural network training. This copying approach allows the system to learn from diverse image compositions without needing actual pre-prepared patches for every scenario, enabling both fast conversion and broad adaptability.
2Measurement precision
If multiple neural networks are maintained for different resolutions, then conversion accuracy is improved, but device complexity increases
Solution Approach 1:
The system employs a single universal neural network model that can handle multiple resolution conversions through one network architecture. This universal approach maintains conversion accuracy across different resolutions while avoiding the complexity of managing multiple separate neural networks, thus resolving the contradiction between precision and device complexity.
Solution Approach 2:
The invention achieves different resolution conversions by changing input parameters (image resolution) to the same neural network model rather than maintaining multiple networks. The neural network is trained to handle variable resolution inputs, allowing accurate conversion across resolutions while keeping the system architecture simple and unified.
Data Source
AI summary
A system previously generates a plurality of neural networks by performing learning based on an image data pair generated based on identical original data and differing in resolution. Then, the system determines one neural network out of the plurality of neural networks based on a resolution of input image data and a resolution of an image to be output, acquires post-conversion image data based on the determined neural network and the input image data, and performs outputting that is based on the post-conversion image data.