Noise Model Training With Noisy Signal Pairs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise reduction techniques require clean and noisy signal pairs for training noise models, which is impractical due to the difficulty in obtaining clean images and the inefficiency of existing denoising methods, leading to high costs and processor-intensive processes.
Innovation Solution
A method for training a noise model and denoiser simultaneously using only pairs of noisy signals captured by a device, eliminating the need for clean images by leveraging a denoiser to produce clean-like images for training, thereby generating a noise model that accurately simulates device-specific noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If clean and noisy signal pairs are used for training noise models, then the noise model can be trained accurately, but it becomes impractical due to the difficulty in obtaining clean images and high processing requirements
Solution Approach 1:
The patent introduces a denoiser as an intermediary component that generates synthetic clean images from noisy images. This denoiser acts as a mediator between the noisy training data and the noise model, enabling the noise model to be trained without requiring actual clean images. The denoiser produces plausible clean image estimates that serve as targets for noise model training, thus resolving the contradiction between training accuracy and data acquisition practicality
Solution Approach 2:
The system uses the noisy images themselves to generate the clean image targets through the denoiser, rather than requiring external clean image sources. The noisy images serve dual purposes: as input data and as the basis for generating their own corresponding clean targets. This self-service approach eliminates the need for separate clean image datasets, making the training process practical while maintaining accuracy
2Measurement precision
If existing denoising methods are used to generate clean images for training, then clean images can be obtained, but the process becomes inefficient and processor-intensive
Solution Approach 1:
The patent combines the denoising process and noise model training into a single integrated workflow. Instead of separately generating clean images using denoising methods and then using them for training, the system merges these operations so that the denoiser and noise model are trained simultaneously using the same noisy image pairs. This integration reduces redundant processing and improves overall training efficiency while maintaining clean image quality
Solution Approach 2:
The denoiser is pre-trained or pre-configured to generate clean image estimates before the noise model training begins. This preliminary action creates a foundation of clean image targets that guide the subsequent noise model training, avoiding the need for iterative or computationally intensive denoising operations during the main training process. The preliminary denoising action enables efficient parallel or sequential training without sacrificing clean image quality
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
A noise model is iteratively trained to simulate introduction of noise by a capture device, by use of a denoiser and a training data set of pairs of noisy signals. First and second noisy signals of each pair are independently sampled by the capture device from source information corresponding to the pair. Each iteration of training obtains first and second denoised signals from respective noisy signals, then optimizes at least one loss function which sums first and second terms to train both the noise model and the denoiser, where the first term is based on the first denoised signal and the second noisy signal, and the second term is based on the second denoised signal and the first noisy signal. By using noisy samples, the complexities of obtaining "clean" signals are avoided. By using "cross-sample" loss functions, convergence on undesired training results is avoided without complex regularization.