Neural Image Denoising With SNR-Adaptive Internal Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks struggle with maintaining noise removal or reduction performance when the amount of noise in the input image differs from that used during training, leading to reduced effectiveness.
Innovation Solution
An image processing system that adjusts internal parameters of a neural network based on signal-to-noise ratio (SNR) data, using functions like Soft-Shrinkage and Hard-Shrinkage activation functions, to adapt to varying noise levels in input images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training images with a single noise magnitude are used, then noise removal performance is optimized for that specific noise level, but performance deteriorates when the actual noise level differs from training noise level
Solution Approach 1:
The patent applies dynamics by making the internal parameters of the neural network adjustable during the processing phase. The system dynamically changes parameters such as the threshold value in the Soft-Shrinkage function based on the actual noise level of the input image, allowing the network to adapt to different noise conditions rather than being fixed to training conditions
Solution Approach 2:
The patent implements parameter changes by modifying internal parameters of the neural network based on the signal-to-noise ratio of the input image. Specifically, the threshold parameter in the activation function is adjusted according to the measured noise level, enabling the network to maintain optimal performance across varying noise magnitudes
2Adaptability or versatility
If multiple training images with different noise amounts are used, then adaptability to varying noise levels is improved, but noise removal performance is reduced compared to single-magnitude training
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing multiple sets of internal parameters corresponding to different noise levels during the training phase. When processing an input image, the system first estimates the noise level and then selects or adjusts the appropriate parameter set, avoiding the need to train on all possible noise levels simultaneously while maintaining performance for each specific level
3Productivity
If internal parameters are fixed after training, then processing speed is maintained, but performance deteriorates when noise levels vary from training conditions
Solution Approach 1:
The patent implements self-service by enabling the neural network to automatically adjust its own internal parameters based on the input image's noise characteristics. The system includes a noise level estimation module that works together with the parameter adjustment mechanism, allowing the network to self-adapt to different noise conditions without external intervention or retraining
Data Source
AI summary
In one embodiment, an image processing system includes a memory and processing circuitry. The memory is configured to store a predetermined program. The processing circuitry is configured, by executing the predetermined program, to perform processing on an input image by exploiting a neural network having an input layer, an output layer, and an intermediate layer provided between the input layer and the output layer, the input image being inputted to the input layer, and adjust an internal parameter based on data related to the input image, while performing the processing on the input image after training of the neural network, the internal parameter being at least one internal parameter of at least one node included in the intermediate layer, and the input parameter having been calculated by the training of the neural network.