✅作者简介：热爱科研的Matlab仿真开发者，修心和技术同步精进，matlab项目合作可私信。
🍎个人主页：Matlab科研工作室
🍊个人信条：格物致知。
⛄ 内容介绍
In order to improve the denoising effect of natural image based-patch prior and effectively remove the noise in the image,this paper proposes Expectation Maximization adaptation learning process by using the statistical characteristics of image blocks to learn image block priors,which generates specific prior by mapping a generic prior to the specified image. Compared with the standard EM algorithm,the proposed method needs less training data,and can be applied to the pre-filtered image in the absence of clean databases. The experimental results show that the proposed algorithm is superior to the existing image denoising algorithms.
⛄ 部分代码
clear;
close all;
addpath('code');
addpath('data/standard_images')
load GSModel_8x8_200_2M_noDC_zeromean.mat
GMM.ncomponents = GS.nmodels;
GMM.mus = GS.means;
GMM.covs = GS.covs;
GMM.weights = GS.mixweights;
clear GS;
x = im2double(imread('House256.png'));
sigmaNoise = 20/255;
y = x + sigmaNoise * randn(size(x)); % noisy test image
%%%% EPLL denoising %%%%
xEPLL = y;
for sigma = sigmaNoise * [1, 1/sqrt(4), 1/sqrt(8), 1/sqrt(16), 1/sqrt(32)]
[xEPLL, psnr_EPLL, ssim_EPLL] = MAP_GMM(x, y, xEPLL, sigmaNoise, sigma, GMM);
end
fprintf('PSNR(EPLL) is:%.2f\n', psnr_EPLL);
fprintf('SSIM(EPLL) is:%.4f\n', ssim_EPLL);
%%%% EM adaptation using EPLL denoised image and MAP denoising with adapted GMM %%%%
xHat = xEPLL;
epsilon = 0.01;
b = randn(size(y));
n = numel(y);
xEPLL1 = y + epsilon*b;
for sigma = sigmaNoise * [1, 1/sqrt(4), 1/sqrt(8), 1/sqrt(16), 1/sqrt(32)]
[xEPLL1, ~, ~] = MAP_GMM(x, y + epsilon*b, xEPLL1, sigmaNoise, sigma, GMM);
end
xHat1 = xEPLL1;
div = (b(:)'*(xHat1(:) - xHat(:))) / (n*epsilon);
beta_opt = (sqrt(mean((y(:) - xHat(:)).^2) - sigmaNoise^2 + 2*sigmaNoise^2*div)) / sigmaNoise;
aGMM = EM_adaptation(GMM, xEPLL, beta_opt * sigmaNoise, 1);
xAdapted_EPLL = y;
for sigma = sigmaNoise * [1, 1/sqrt(4), 1/sqrt(8), 1/sqrt(16), 1/sqrt(32)]
[xAdapted_EPLL, psnr_adapted, ssim_adapted] = MAP_GMM(x, y, xAdapted_EPLL, sigmaNoise, sigma, aGMM);
end
fprintf('PSNR(adapted by EPLL image) is:%.2f\n', psnr_adapted);
fprintf('SSIM(adapted by EPLL image) is:%.4f\n', ssim_adapted);
figure
subplot(131)
imshow(x);
title('原图')
subplot(132)
imshow(y);
title('加噪图')
subplot(133)
imshow(xAdapted_EPLL);
title(['去噪图,PSNR=',num2str(psnr_adapted)])
return
⛄ 运行结果
⛄ 参考文献
[1] Lei Y . Baesd-patch Daptive Image Denoising with EM Algorithm[J]. Science Mosaic, 2017.
