Neural Network Noise Removal via Simulated Signal Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current noise removal systems for radar images, such as those using neural networks, face challenges in training due to the absence of clean images and reliance on noisy data, which can lead to errors and inefficiencies.
Innovation Solution
Training a neural network to remove noise by comparing the removed noise to simulated noise, using a noise simulator to generate noise characteristics matching the data source, and employing a comparison value to improve the network's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If clean images are used for training neural networks, then training accuracy is improved, but clean images do not exist for many radar systems because noise is present in the received data
Solution Approach 1:
The patent creates synthetic clean images by copying the structure and content of noisy images while removing noise through simulation. A noise simulator generates expected noise patterns that are subtracted from noisy images to produce synthetic clean training images, allowing the neural network to be trained on realistic clean data that would otherwise not exist.
Solution Approach 2:
The patent introduces a noise simulator as an intermediary component that generates expected noise patterns. This intermediary enables the comparison between actual noise in images and simulated noise, facilitating the creation of training data and providing a basis for evaluating noise removal performance without requiring genuine clean images.
2Quantity of substance
If multiple noisy images are used for training, then more training data is available, but the system may train on spikes in the noise data which results in errors
Solution Approach 1:
Instead of copying multiple noisy images which would propagate noise artifacts, the patent copies the underlying signal structure while generating clean versions through noise subtraction. Synthetic clean images are created by removing simulated noise patterns from noisy images, providing abundant training data without copying noise errors.
Solution Approach 2:
The system uses the noisy images themselves to generate their own clean counterparts through noise simulation and subtraction. The noise simulator generates expected noise patterns that are removed from the noisy images, allowing the data to serve its own training purpose without requiring external clean references or risking propagation of noise spikes.
3Productivity
If noise removal is performed by traditional filtering techniques, then processing speed is maintained, but noise removal effectiveness is reduced compared to neural network approaches
Solution Approach 1:
The patent performs preliminary noise simulation and characterization during the training phase, allowing the neural network to learn optimal noise removal strategies in advance. By pre-generating synthetic clean images and training the network beforehand, the system achieves fast real-time processing while maintaining high noise removal effectiveness through the pre-learned neural network model.
Data Source
AI summary
A neural network is used to remove noise from a data signal. The noise removed by the neural network is compared to simulated noise that represents noise expected to be present in the data signal. The results of the comparison are used to train the neural network and improve its ability to remove noise from the data signal.