Image denoising neural network architecture and method of training the same
AssigneeSamsung Electronics Co., Ltd.
InventorsMostafa El-Khamy, Igor Fedorov, Jungwon Lee
Life of this patent
Estimating the term…
Abstract
An image denoising neural network training architecture includes an image denoising neural network and a clean data neural network, and the image denoising neural network and clean data neural network share information between each other.
Claims22
An image denoising neural network training architecture comprising:
- an image denoising neural network;
- and
- a clean data neural network,
- wherein the image denoising neural network and clean data neural network are configured to share information between each other.
A method of training an image denoising neural network, the image denoising neural network comprising a variance-stabilizing transform network, an inverse variance-stabilizing transform network, and a denoising network, the method comprising:
- jointly training the variance-stabilizing transform network and the inverse variance-stabilizing transform network;
- and
- training the denoising network.
An image denoising neural network training architecture comprising:
- an image denoising neural network, the image denoising network comprising:
- a variance-stabilizing transform (VST) network comprising a plurality of filtering layers;
- an inverse variance-stabilizing transform (IVST) network comprising a plurality of filtering layers;
- and
- a denoising network comprising a plurality of filtering layers;
- and
- a clean data network, the clean data network comprising a plurality of filtering layers,
- wherein a plurality of the filtering layers of the clean data network are configured to communicate with a corresponding plurality of the filtering layers of the denoising network.
Dependencies are inferred from the claim text.