Code-Level Super-Resolution Using Encoded Image Codes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution imaging techniques are dependent on the computing resources of user equipment, making them inefficient and resource-intensive.
Innovation Solution
A method for training a super-resolution model to perform imaging at the code level by encoding and decoding encoded codes, using a neural network model to predict high-resolution encoded codes from low-resolution encoded codes, and training the model based on correct-answer and inference sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If super-resolution imaging is performed at the picture level using existing models, then high-resolution images can be generated, but the computing resources of user equipment are heavily consumed
Solution Approach 1:
The patent segments the super-resolution process into two distinct stages: (1) a training stage performed on a server using picture-level data, and (2) an inference stage performed on user equipment using code-level data. This segmentation allows the computationally intensive training to be done once on powerful servers, while the lightweight inference runs efficiently on user devices, thereby reducing ongoing computing resource consumption.
Solution Approach 2:
The patent performs preliminary action by pre-training the super-resolution model on the server using high-quality picture-level training data before deployment. The trained model parameters are then deployed to user equipment, eliminating the need for users to perform resource-intensive training locally. This preliminary training phase transfers the computational burden from end-user devices to centralized servers.
2Measurement precision
If super-resolution processing is performed on full-resolution pictures, then accurate high-resolution output is achieved, but data size and transmission requirements increase
Solution Approach 1:
The patent inverts the conventional approach by working in the code domain rather than the picture domain. Instead of taking low-resolution pictures and upscaling them, the system takes compressed codes as input, processes them through the trained model to generate enhanced codes, and then decodes to produce high-resolution images. This inversion allows processing of compact representations rather than large image data.
Solution Approach 2:
The patent introduces compressed codes as an intermediary representation between the original image and the super-resolution output. The training process learns to map between code representations, and the inference process uses this learned mapping to enhance images. This intermediary code domain serves as a compact bridge that reduces data transmission requirements while preserving essential image information.
Data Source
AI summary
The present disclosure provides a method for code-level super-resolution imaging and a method of training a super-resolution model therefor. According to some embodiments, the present disclosure provides a method of training a super-resolution model based on the encoded codes of an inputted picture and a method for code-level super-resolution imaging by using a pre-trained super-resolution model.