Blind Image Denoising With Training-Distribution Noise Matching
Solution Overview
Problem
Conventional blind denoisers trained using Gaussian noise are ineffective in removing real-world noise that does not follow a Gaussian distribution, leading to unsatisfactory image denoising results.
Innovation Solution
Apply synthetic noise based on a matching distribution to an input image to create a noise-added image, which is then processed by a blind denoiser to improve denoising effectiveness, and optionally combine multiple output images to enhance the result.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a blind denoiser is trained using Gaussian noise, then it can effectively remove Gaussian-distributed noise, but it fails to remove real-world noise that does not follow a Gaussian distribution
Solution Approach 1:
The patent changes the noise distribution parameter from Gaussian to a distribution that matches the training data (e.g., uniform distribution or empirical noise distribution from real images). By adjusting the noise distribution parameter to match actual training conditions, the denoiser achieves both reliability in removing specific noise types and adaptability through configurable distribution selection.
Solution Approach 2:
The patent applies preliminary noise addition to the input image before processing it through the denoiser. This preliminary action transforms the input image by adding synthetic noise that matches the training distribution, enabling the denoiser to effectively process images with noise distributions different from Gaussian, thereby improving both denoising effectiveness and adaptability.
2Reliability
If synthetic noise is added to match training data distribution, then denoising performance improves, but image processing complexity increases
Solution Approach 1:
The patent performs preliminary noise addition with a distribution matching the training data before the denoising operation. This simple preprocessing step significantly improves denoising performance without requiring complex model architectures or multiple processing stages, thus achieving high reliability with minimal increase in processing complexity.
Solution Approach 2:
The patent introduces synthetic noise as an intermediary element between the input image and the denoiser. This intermediary noise, configured to match the training distribution, acts as a bridge that enables the denoiser to effectively process images with various noise types, improving performance while keeping the overall system relatively simple.
3Loss of time
If the denoiser is trained on synthetic Gaussian noise, then training is straightforward and fast, but the denoiser cannot handle real-world noise distributions
Solution Approach 1:
Instead of retraining the denoiser on real-world noisy images (which would be time-consuming), the patent applies preliminary noise addition during inference that matches the training distribution. This approach maintains the fast training advantage while achieving reliable real-world denoising performance without requiring additional training time.
Solution Approach 2:
The patent creates a copy of the training noise distribution and applies it to the input image during processing. By copying the statistical characteristics of the training noise (rather than retraining on real noise), the system maintains fast training while achieving reliable real-world denoising capability through distribution matching.
Data Source
AI summary
A device includes a memory configured to store an input image. The device also includes one or more processors configured to apply synthetic noise to the input image to generate a noise-added image and to apply a denoiser to the noise-added image to generate an output image that has less noise than the input image.