Regional Generative and CNN Image Processing to Reduce Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods using machine learning models struggle to effectively sharpen blur on focal planes and shape defocus blur on non-focal planes without generating artificial structures, particularly in luminance-saturated areas and areas with low optical performance.
Innovation Solution
A method involving a combination of convolutional neural networks (CNN) for blur sharpening on focal planes and generative models like diffusion models for defocus blur shaping on non-focal planes, with weights assigned based on segmentation maps, optical performance, and saturated areas to control blur sharpening effects and prevent artificial structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If generative models are used for image deblurring and enhancement, then performance in regression tasks is improved, but artificial structures are generated in luminance-saturated areas and areas with low optical performance
Solution Approach 1:
The image is divided into multiple areas based on optical performance characteristics and luminance saturation. Different machine learning models are selectively applied to different segments: generative models for areas with high optical performance and non-generative models for areas with low optical performance or high saturation, thereby reducing artificial structure generation while maintaining enhancement accuracy.
Solution Approach 2:
Different regions of the image are processed with different models according to their local characteristics. Areas with low optical performance or high luminance saturation use non-generative models to avoid artificial structures, while areas with high optical performance use generative models for enhanced accuracy, achieving local optimization of quality versus artifact generation.
2Manufacturing precision
If multiple machine learning models are combined with different weights for different areas, then image processing accuracy is improved, but system complexity increases
Solution Approach 1:
The image domain is segmented into multiple areas based on optical performance and luminance saturation characteristics. A plurality of machine learning models are prepared with different weight values, and these models are selectively applied to different segmented areas. This segmentation approach enables accurate image processing by matching appropriate models to specific regions while managing system complexity through structured organization of models and areas.
Data Source
AI summary
An image processing method includes a step of generating an estimated image from an input image by using a plurality of machine learning models including a generative model and a non-generative model. In the step, the estimated image is generated by assigning different weights to output of the generative model and output of the non-generative model for each of a plurality of areas of the input image based on information regarding the input image.