Denoiser Training via Pseudo-Labels and Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing noise reduction technologies, such as Discrete Universal DEnoiser (DUDE) and deep learning methods, face challenges in achieving optimal performance due to sensitivity to hyperparameters and the need for large amounts of training data, particularly in selecting an appropriate sliding window size and securing clean data for supervised learning.
Innovation Solution
A denoising device and method that utilize pseudo-labels derived from context and estimated losses to train deep learning algorithms, allowing for the optimization of denoiser parameters and sliding window size without requiring clean data, by using a combination of Discrete Universal DEnoiser and deep neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with massive training data is used, then learning accuracy is improved, but data acquisition cost and complexity increase significantly
Solution Approach 1:
The system generates its own training data by using the denoiser to create pseudo-clean samples from noisy data. The denoiser processes noisy images to generate reconstructed images, which then serve as training pairs (noisy image, reconstructed image) for supervised learning, eliminating the need for external clean data sources
Solution Approach 2:
The denoiser is pre-trained or configured to generate reasonable reconstructions before the main learning process. This preliminary denoising step creates initial training data that can be used to train the learning model, allowing the system to bootstrap the learning process without requiring pre-existing clean training datasets
2Reliability
If sliding window size is increased to improve context representation, then noise reduction performance is improved, but computational complexity and parameter selection difficulty increase
Solution Approach 1:
The system employs multiple denoisers with different sliding window sizes (k=1, 2, 3) rather than a single fixed size. The learning model dynamically selects and combines results from these multiple denoisers based on the specific input context, allowing adaptive optimization without requiring manual selection of a single optimal parameter
Solution Approach 2:
Multiple denoisers with different sliding window sizes are merged into a unified learning framework. The system combines the outputs of denoisers with k=1, k=2, and k=3 through the learning model to achieve superior noise reduction performance, leveraging the strengths of each window size without requiring separate systems
3Reliability
If multiple denoisers with different sliding window sizes are trained, then noise reduction performance is improved, but training data requirements and processing time increase
Solution Approach 1:
The training processes for multiple denoisers are merged into a single unified learning operation. The learning model is trained once on datasets generated by all denoisers simultaneously, rather than training each denoiser separately. This allows parallel processing and shared computational resources, reducing total training time while maintaining the performance benefits of multiple denoisers
Data Source
AI summary
Disclosed herein is a denoising device including a deriving part configured to, when corrupted noise data corrupted due to noises is received from source data, derive an estimated loss which is estimated when each symbol within noise data is reconstructed to the source data based on a predefined noise occurrence probability, a processor to process training of a defined learning model by including parameters related with the reconstruction of the source data from the noise data based on context composed of a sequence of neighbored symbols based on each symbol within the noise data and pseudo-training data using the estimated loss corresponding to the context, and an output part to output reconstructed data in which each symbol within the noise data is reconstructed to a symbol of the source data through a denoiser formed based on a result of the training processing.