Image Processing Model Training With Conversion-Frequency Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image quality improvement technologies using machine learning require complex procedures and high calculation costs for multi-task learning to control the balance between sharpness and image quality effects.
Innovation Solution
A learning apparatus that controls the frequency and level of conversion processes on input images, training multiple models with different image quality characteristics, allowing for the application of a designated model based on the desired image quality characteristic.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multi-task learning is performed with a large-scale model to control the balance between sharpness and image quality improvement effect, then the image quality characteristics can be adjusted, but the calculation cost of learning increases and the procedure becomes complicated
Solution Approach 1:
The patent segments the learning process by training multiple specialized models for different image quality characteristics (noise reduction, super-resolution, etc.) rather than using a single multi-task model. Each model is trained independently on specific degradation types, simplifying the overall learning procedure while maintaining versatility through model selection and combination.
Solution Approach 2:
The patent creates a universal framework where multiple single-function models can be combined to achieve multi-task capabilities. By training separate models for different image quality improvements and selecting/applying appropriate models based on input image characteristics, the system achieves versatility without the complexity of multi-task learning in a single model.
2Adaptability or versatility
If multi-task learning is performed with a large-scale model to control the balance between sharpness and image quality improvement effect, then the image quality characteristics can be adjusted, but the calculation cost of learning increases
Solution Approach 1:
The patent divides the learning task into multiple specialized single-task models, each optimized for a specific image quality improvement. This segmentation allows for more efficient training of individual models compared to a large-scale multi-task model, reducing overall calculation costs while maintaining the ability to adjust image quality characteristics through model selection.
Solution Approach 2:
The patent creates multiple copies of base models with different specializations (noise reduction, super-resolution, etc.) rather than training one large multi-task model. This approach allows parallel training of smaller, more efficient models that can be selectively applied, reducing the computational burden compared to training a single comprehensive model.
3Manufacturing precision
If the strength of image quality improvement effect is increased, then noise reduction and other quality improvements are enhanced, but the sharpness of the output image decreases
Solution Approach 1:
The patent implements dynamic model selection and parameter adjustment based on the input image characteristics and desired output properties. By selecting appropriate models and adjusting their parameters dynamically, the system can optimize the balance between image quality improvement effect and sharpness preservation for each specific processing scenario.
Solution Approach 2:
The patent changes parameters such as the strength of degradation simulation, learning rates, and model selection criteria to control the trade-off between image quality improvement and sharpness. By adjusting these parameters based on the specific task requirements, the system can achieve optimal performance for different image quality characteristics.
Data Source
AI summary
A learning apparatus includes a conversion controller that controls an occurrence frequency depending on a level of a first conversion process as a subject included in one or more conversion processes to be performed on a first image for learning; a model processor that inputs a second image obtained by performing the conversion processes on the first image to a model trained based on machine learning and causes the model to output a third image obtained by performing image processing on the second image; and a loss calculator that calculates a loss between the third image and the first image. A plurality of models having different image quality characteristics of the image processing are generated for each combination of the level and the occurrence frequency. A model corresponding to a designated image quality characteristic included in the plurality of models is applied to an image processing apparatus.