Image denoising method
There is provided an image denoising method including extracting a noise patch from a noisy image, outputting a noise parameter by inputting the noise patch to a noise parameter estimation network (NPE-net), generating imitated virtual noise based on the output noise parameter, generating a noisier image by adding the imitated virtual noise to the noisy image, training a denoise deep learning model by inputting the noisy image and the noisier image as a pair to the denoise deep learning model, inputting the noisy image to the trained denoise deep learning model, and outputting a denoise image obtained by removing noise from the noisy image by the trained denoise deep learning model.
1 . An image denoising method comprising:
extracting a noise patch from a noisy image;
outputting a noise parameter by inputting the noise patch to a noise parameter estimation network (NPE-net);
generating imitated virtual noise based on the output noise parameter;
generating a noisier image by adding the imitated virtual noise to the noisy image;
training a denoise deep learning model by inputting the noisy image and the noisier image as a pair to the denoise deep learning model;
inputting the noisy image to the trained denoise deep learning model; and
outputting a denoise image obtained by removing noise from the noisy image by the trained denoise deep learning model.
2 . The image denoising method of claim 1 , wherein the noisy image comprises a scanning electron microscope (SEM) image obtained with an SEM for a semiconductor process result.
3 . The image denoising method of claim 2 , wherein the noise patch targets only a noise portion that does not include a pattern shape of the semiconductor process result in the semiconductor process result.
4 . The image denoising method of claim 1 , wherein a pair composed by taking a plurality of pieces of learning noise data as inputs and noise parameters of the plurality of pieces of learning noise data as outputs is input to the NPE-net to train the NPE-net.
5 . The image denoising method of claim 4 , wherein each of the plurality of pieces of learning noise data includes virtual learning noise data having a random noise parameter.
6 . The image denoising method of claim 1 , wherein the NPE-net assumes that noise included in the noise patch is Gaussian noise following a Gaussian distribution, and the noise parameter includes a noise average and a noise standard deviation.
7 . The image denoising method of claim 1 , wherein the denoise deep learning model is trained to receive the noisier image and output the noisy image.
8 . The image denoising method of claim 1 , wherein the outputting of the denoise image includes calculating noise M obtained by subtracting the noisy image from the noisier image input to the trained denoise deep learning model, and calculating the denoise image by subtracting twice the noise M from the noisier image.
9 . The image denoising method of claim 1 , wherein the denoise deep learning model is learned through an L2 loss function.
10 . The image denoising method of claim 1 , further comprising, after the outputting of the denoise image, inputting the denoise image into an optical proximity correction (OPC) model.
11 . The image denoising method of claim 1 , further comprising:
after the outputting of the denoise image, determining whether noise included in the denoise image is sufficiently remove;
subsequent to determining that the noise included in the denoise image is sufficiently removed, terminating the image denoising method; and
subsequent to determining that the noise included in the denoise image is not sufficiently removed, performing the extracting of the noise patch to the outputting of the denoise image again by using the denoise image as the noisy image.
12 . The image denoising method of claim 1 , further comprising:
after the outputting of the denoise image, determining whether noise included in the denoise image is sufficiently removed;
subsequent to determining that the noise included in the denoise image is sufficiently removed, terminating the image denoising method; and
subsequent to determining that the noise included in the denoise image is not sufficiently removed, performing the generating of the noisier image to the outputting of the denoise image again by using the denoise image as the noisy image.
13 . A scanning electron microscope (SEM) image denoising method comprising:
training a noise parameter estimation network (NPE-net) with noise data for learning;
extracting a noise patch from a non-pattern region included in an SEM image obtained with an SEM for a semiconductor process result;
outputting a noise parameter by inputting the noise patch to the NPE-net;
generating imitated virtual noise based on the output noise parameter;
generating a noisier image by adding the imitated virtual noise to the SEM image;
training a denoise deep learning model by inputting the SEM image and the noisier image as a pair to the denoise deep learning model;
inputting the SEM image to the trained denoise deep learning model; and
outputting a denoise image obtained by removing noise from the SEM image by the trained denoise deep learning model.
14 . The SEM image denoising method of claim 13 , wherein the NPE-net is trained with a pair composed by taking, as an input, the noise data for learning composed of virtual noise data for learning having a random standard deviation of Gaussian distribution noise and taking, as an output, a standard deviation of the Gaussian distribution noise corresponding to the noise data for learning.
15 . The SEM image denoising method of claim 14 , wherein the outputting of the denoise image includes calculating noise M obtained by subtracting the SEM image from the noisier image input to the trained denoise deep learning model, and outputting the denoise image by subtracting twice the noise M from the noisier image,
wherein the noise M is Gaussian distribution noise.
16 . The SEM image denoising method of claim 15 , wherein a Gaussian distribution average of the noise M is 0.
17 . The SEM image denoising method of claim 13 , wherein the noise patch includes a plurality of noise patches extracted from the non-pattern region, and subsequent operations are performed in parallel or individually by using the plurality of the noise patches to output a plurality of denoise images.
18 . The SEM image denoising method of claim 13 , further comprising:
after the outputting of the denoise image, determining whether noise included in the denoise image is sufficiently removed;
subsequent to determining that the noise included in the denoise image is sufficiently removed, terminating the SEM image denoising method; and
subsequent to determining that noise included in the denoise image is not sufficiently removed, performing the extracting of the noise patch to the outputting of the denoise again by using the denoise image as the SEM image.
19 . A scanning electron microscope (SEM) image denoising method comprising:
training a noise parameter estimation network (NPE-net) with a pair composed by taking, as an input, noise data for learning composed of virtual noise data for learning having a random standard deviation of Gaussian distribution noise and taking, as an output, a standard deviation of the Gaussian distribution noise corresponding to the noise data for learning;
extracting a noise patch from a non-pattern region included in an SEM image obtained with an SEM for a semiconductor process result;
outputting a noise parameter by inputting the noise patch to the NPE-net;
generating imitated virtual noise based on the output noise parameter;
generating a noisier image by adding the imitated virtual noise to the SEM image;
training a denoise deep learning model by inputting the SEM image and the noisier image as a pair to the denoise deep learning model, wherein the denoise deep learning model is trained through an L2 loss function to receive the noisier image and output the SEM image;
inputting the SEM image to the trained denoise deep learning model; and
outputting a denoise image obtained by removing noise from the SEM image by the trained denoise deep learning model, wherein noise M is calculated by subtracting the SEM image from the noisier image and the denoise image is calculated by subtracting twice the noise M from the noisier image.
20 . The SEM image denoising method of claim 19 , further comprising:
after the outputting of the denoise image, determining whether noise included in the denoise image is sufficiently removed;
subsequent to determining that the noise included in the denoise image is sufficiently removed, terminating the SEM image denoising method; and
subsequent to determining that the noise included in the denoise image is not sufficiently removed, performing the extracting of the noise patch to the outputting of the denoise image again by using the denoise image as the SEM image,
wherein the noise patch includes a plurality of noise patches extracted from the non-pattern region, and subsequent operations are performed in parallel or individually by using the plurality of the noise patches to output a plurality of denoise images.