CPC G06T 5/002 (2013.01) [G06N 3/047 (2023.01); G06N 3/048 (2023.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01)] | 20 Claims |
1. A method for denoising an image, the method comprising:
training, utilizing one or more processors, a fully convolutional neural network (FCN) by:
generating an nth training image of a plurality of training images, the nth training image comprising a plurality of training channels and a training array In of size Nh×Nw×C, where:
Nh is a height of the nth training image,
Nw is a width of the nth training image,
C is a number of the plurality of training channels,
1≤n≤N, and
N is a number of the plurality of training images;
initializing the FCN with a plurality of initial weights; and
repeating a first iterative process until a first termination condition is satisfied, comprising:
extracting an nth denoised training image from an output of the FCN by applying the FCN on the nth training image, the nth denoised training image comprising a denoised array În of size Nh×Nw×C;
generating a plurality of updated weights by minimizing a loss function comprising Σn=1N∥In−În| where |.| is an L1 norm, each updated weight of the plurality of updated weights associated with a respective initial weight of the plurality of initial weights; and
replacing the plurality of initial weights with the plurality of updated weights; and
generating, utilizing the one or more processors, a reconstructed image by applying the FCN on the image.
|