Neural Network Parameter Update During High-Resolution Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep-learning based image processing systems face challenges in updating parameters of neural network layers without stopping calculations, leading to artifacts in output images due to the large number of parameters and complex image processing requirements.
Innovation Solution
An electronic apparatus and method that allows for real-time parameter updates in neural network systems by operating in two modes: one where the output image is processed using an updated artificial intelligence model, and another where the output is based solely on interpolated images, enabling efficient parameter updates without stopping calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parameters of neural network layers are updated during high-resolution image generation, then image processing efficiency is improved, but artifacts appear in output images due to the large number of parameters
Solution Approach 1:
The patent segments the parameter update process into two distinct modes: a first mode where parameters are updated and a second mode where parameter updates are suspended. This segmentation allows the system to switch between updating parameters and maintaining stable output, thereby resolving the contradiction between improving processing efficiency and maintaining output image quality.
Solution Approach 2:
The patent implements periodic switching between the first mode (parameter update mode) and the second mode (stable output mode). By periodically alternating between parameter updates and stable generation, the system achieves both efficient processing through updates and high-quality output during stable periods, thus resolving the technical contradiction.
2Loss of time
If parameters are updated without stopping calculation, then processing continuity is maintained, but artifacts appear in the output image
Solution Approach 1:
The patent introduces dynamic mode switching capability that allows the system to adapt between two operational states: updating parameters while maintaining calculation continuity in the first mode, and suspending parameter updates to ensure output quality in the second mode. This dynamic adjustment resolves the contradiction between calculation continuity and output image quality.
Solution Approach 2:
The patent changes the operational parameter of the neural network system by switching between two modes: in the first mode, parameter updates are enabled to maintain calculation continuity; in the second mode, parameter updates are disabled to prevent artifacts. This parameter change approach allows the system to balance calculation continuity with output quality.
3Manufacturing precision
If deep-learning based processing is used for high-resolution images, then image quality is improved, but the large number of parameters makes parameter updates difficult
Solution Approach 1:
The patent segments the complex parameter update problem by dividing it into two manageable modes: a first mode that handles parameter updates and a second mode that handles stable high-quality generation. This segmentation simplifies the overall complexity by allowing each mode to focus on its specific function, thus resolving the contradiction between image quality and parameter update complexity.
Solution Approach 2:
The patent ensures continuous useful action by implementing a switching mechanism between two modes: the first mode performs parameter updates to improve the model, and the second mode performs stable high-quality image generation. This continuous alternation between updating and generating maintains both model improvement and high-quality output, resolving the complexity issue.
Data Source
AI summary
Disclosed is an electronic apparatus. The electronic apparatus includes: a memory configured to store information regarding an artificial intelligence model including a plurality of layers; and a processor configured to perform interpolation processing on an input image and to process the interpolated image using the artificial intelligence model to obtain an output image, wherein the processor is configured to be operated in a first mode or a second mode based on an update of parameters used in at least one of the plurality of layers being required, the first mode including a mode in which the output image is obtained based on an image processed using the artificial intelligence model in which the parameters are updated and based on the interpolated image, and the second mode includes a mode in which the output image is obtained based on the interpolated image.