Single Converter Denoising Neural Network for Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional denoising techniques require multiple neural networks to handle various types of noise, leading to high computational demands and limited practical implementation, especially in edge devices with limited processing capacity.
Innovation Solution
A single machine learning model is trained using sensor data with noise and noise-free data from similar scenes to reduce noise across multiple types, utilizing Generative Adversarial Networks and feature extractors to update parameters and improve denoising performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple neural networks are used to handle various types of noise, then denoising performance for different noise types is improved, but computational load increases
Solution Approach 1:
The patent combines multiple specialized neural networks into a single unified denoiser that can handle various noise types. This is achieved by training a single network with diverse noise data during the learning phase, allowing it to generalize across different noise conditions without requiring separate models for each noise type, thereby reducing computational overhead while maintaining denoising effectiveness
Solution Approach 2:
The denoiser is designed as a universal model capable of processing multiple noise types simultaneously. By incorporating multi-functionality into a single network architecture, the system can adapt to different noise conditions (e.g., shot noise, Gaussian noise, Poisson noise) without requiring multiple specialized networks, thus resolving the contradiction between performance and computational efficiency
2Measurement precision
If multiple specialized models are deployed for different noise types, then denoising accuracy is improved, but device complexity increases
Solution Approach 1:
The patent merges multiple specialized denoising models into a single integrated model. During training, the network is exposed to various noise types and learns to distinguish and process them appropriately. This consolidation reduces device complexity by eliminating the need to manage, deploy, and coordinate multiple separate models while preserving the ability to achieve high denoising accuracy across different noise conditions
3Use of energy by moving object
If a single converter is used for all noise types, then computational load is reduced, but denoising performance across various noise types deteriorates
Solution Approach 1:
The patent applies preliminary action by extensively training the single converter model with diverse noise data before deployment. During the training phase, the network learns to recognize and handle various noise types through exposure to synthesized and real noisy images. This preliminary learning equips the model with the capability to maintain high denoising performance across different noise types while keeping the runtime computational load low, as no complex model selection or switching is needed during actual denoising operations
Data Source
AI summary
An information processing method performed using a computer includes: obtaining first sensor data (SD) that is SD of a scene and includes noise; inputting the first SD to a single converter, and obtaining second SD outputted from the single converter as a result of denoising performed on the first SD by the single converter; obtaining third SD that is SD of a scene identical or corresponding to the scene, does not include the noise, and is different from the second SD; obtaining feature information of the second SD and feature information of the third SD, based on the second SD and the third SD, respectively; and training the single converter by machine learning using the second SD and the feature information of the second SD as converted data, and using the third SD and the feature information of the third SD as reference data corresponding to the converted data.