Image Processing Device Adaptive Model Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in efficiently processing and adapting to varying image qualities, especially when characteristics of an image to be processed differ from those assumed during training, leading to domain gap issues and high computational complexity.
Innovation Solution
An image processing device that stores a cumulative quality of content based on viewing frequency, determines a model storing condition, obtains a reference model, and generates a target model by training the reference model using training data corresponding to the quality of a specific image, thereby reducing computational complexity and improving adaptability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ZSSR creates a new DB adaptively for each input image from scratch and trains a model by using the new DB, then image quality enhancement is achieved, but computational complexity of learning becomes excessively high
Solution Approach 1:
The patent pre-creates a database of degraded images and their corresponding high-quality images before actual processing. This preliminary database construction allows the system to avoid training from scratch for each new image, significantly reducing computational complexity while maintaining image quality enhancement effectiveness.
Solution Approach 2:
The patent uses a pre-trained model that has learned from a comprehensive database of image degradations. Instead of training a new model for each input image, the system copies and adapts the pre-trained model's knowledge, reducing the computational burden from training complexity to inference and adaptation complexity.
2Device complexity
If a single meta model is used to handle all input image features, then computational complexity is reduced, but performance is limited because all features are included in the single model
Solution Approach 1:
The patent segments the single meta model into multiple specialized models, each trained on specific subsets of image features and degradation types. This segmentation allows each model to specialize in handling particular characteristics (e.g., compression artifacts, blur, noise), improving overall performance while maintaining manageable computational complexity through selective model deployment.
Solution Approach 2:
The patent applies local quality by creating different model variants tailored to specific local characteristics of input images. Instead of a uniform model handling all features, the system selects and applies model variants that are optimized for the specific degradation characteristics present in each input image, improving processing performance for each local feature type.
3Adaptability or versatility
If multiple meta models are used to select the suitable model for each input image, then adaptability to input image characteristics is improved, but the ability to train models by predicting all real input images is lost
Solution Approach 1:
The patent implements a feedback mechanism where the system continuously monitors input image characteristics and adjusts model selection accordingly. The feedback loop allows the system to learn from actual input images and refine its model selection strategy, maintaining adaptability while preserving the ability to train models effectively through iterative improvement and validation.
Data Source
AI summary
An image processing device is configured to store a cumulative quality of content comprising a plurality of input images based on a viewing frequency of the content, determine a model storing condition based on the viewing frequency and the cumulative quality, obtain a reference model corresponding to the model storing condition, store the reference model in a memory, and generate a target model corresponding to a first image by training the reference model stored in the memory by using training data corresponding to a quality of the first image.