Image Denoising via Specialized Models and Generalizer Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image denoising techniques face a tradeoff between the high denoising quality of specialized denoisers and the robustness of blind denoisers, with few methods effectively balancing both aspects.
Innovation Solution
A computer-implemented method involving multiple specialized denoisers trained for specific noise ranges and a generalizer that generates per-pixel denoising kernels for arbitrary noise parameters, allowing for improved denoising quality and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If specialized denoisers are trained on specific noise parameters, then denoising quality is improved for those parameters, but robustness to different noise parameters deteriorates
Solution Approach 1:
The patent divides the denoising task into multiple specialized denoisers, each trained on specific noise parameter ranges. Instead of using a single general-purpose denoiser, the system segments the problem space and creates specialized models for different noise conditions, allowing each segment to excel at its specific task while maintaining overall system versatility.
Solution Approach 2:
The patent creates a universal denoising system that can handle multiple noise parameter ranges by combining multiple specialized denoisers. The system achieves multi-functionality by integrating several specialized models that collectively cover a broad range of noise conditions, making the overall system adaptable to diverse scenarios.
2Adaptability or versatility
If blind denoisers are trained on diverse noise parameters, then robustness is improved, but overall denoising quality deteriorates
Solution Approach 1:
Rather than using a single blind denoiser trained on diverse data, the patent segments the diverse noise parameter space into multiple specialized denoisers. Each denoiser is trained on a specific subset of noise parameters, allowing them to achieve high quality for their designated range while the collection as a whole maintains robustness across all ranges.
Solution Approach 2:
The patent applies local quality by making different parts of the system (different specialized denoisers) have different expertise tailored to specific noise conditions. Each denoiser is optimized for its local domain of noise parameters, achieving high denoising quality locally while the system as a whole provides global robustness.
3Adaptability or versatility
If multiple specialized denoisers are used to cover different noise ranges, then robustness is improved, but device complexity increases
Solution Approach 1:
The patent introduces dynamics by implementing a selection mechanism that adaptively chooses which specialized denoiser to use based on the input image's noise characteristics. This dynamic approach allows the system to maintain robustness across different noise conditions while managing complexity through intelligent routing rather than simultaneously deploying all denoisers for every input.
Solution Approach 2:
The patent introduces an intermediary component that acts as a mediator between the input image and the multiple specialized denoisers. This intermediary analyzes the noise parameters and routes the image to the most appropriate denoiser, reducing the effective complexity by avoiding the need for all denoisers to process every image.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
Techniques are disclosed for training and applying a denoising model. The denoising model includes multiple specialized denoisers and a generalizer, each of which is a machine learning model. The specialized denoisers are trained to denoise images associated with specific ranges of noise parameters. The generalizer is trained to generate per-pixel denoising kernels for denoising images associated with arbitrary noise parameters using outputs of the specialized denoisers. Subsequent to training, a noisy image, such as a live-action image or a rendered image, can be denoised by inputting the noisy image into the specialized denoisers to obtain intermediate denoised images that are then input, along with the noisy image, into the generalizer to obtain per-pixel denoising kernels, which can be normalized and applied to denoise the noisy image.