1 算法介绍
传统的超分辨重建算法往往采用梯度下降法进行求解,迭代时步长往往通过经验确定。而且不同的图像的最优步长往往不相同。步长过大会导致发散,步长过小会导致收敛缓慢。本算法基于对正则化超分辨重建算法实现的基础上,对步长的选取进行了优化,推导出了每次迭代时的最优步长大小,并将其自适应化,改进了超分辨算法的收敛性,从而能够在更短的时间内取得更加精确的重建结果。相关具体内容请参考对应的论文:Yingqian Wang, Jungang Yang, Chao Xiao, and Wei An, "Fast convergence strategy for multi-image superresolution via adaptive line search," IEEE Access, vol. 6, no. 1, pp. 9129-9139.
2 部分代码
%% BIBTEX TO CITE:
% @article{wang2018fast,
% title={Fast Convergence Strategy for Multi-Image Superresolution via Adaptive Line Search},
% author={Wang, Yingqian and Yang, Jungang and Xiao, Chao and An, Wei},
% journal={IEEE Access},
% volume={6},
% pages={9129--9139},
% year={2018},
% publisher={IEEE}
% }
%% %%%%%%%%Copyright (c) 2018, All rights reserved.%%%%%%%%%%%%%%%%%%%%%%
clear all
clc
filename = 'Set';
files = dir(fullfile( filename,'*.bmp'));
file_num = 2; % different number corresponds to defferent test images in 'Set'
reg_term = 1; %regularization term: 1-BTV, 2-Tikhonov
Image =imread([filename,'\',files(file_num).name]);
SZ = size(size(Image));
if (SZ(2)==2) % turn grayscale image to RGB image
for qw = 1:3
IMAGE (:,:,qw) = Image;
end
else
IMAGE = Image;
end
%% Image Degradation
D = [1,1;-2,1;-1,-3;3,-2]; % Shearing shift
Gau = fspecial( 'gaussian', [3 3], 1); % Gaussian bluring kernel
spf = 2; % sampling factor
sigma2 = 1; % variation of noise
LR = ImDegrate(IMAGE,D,Gau,spf,sigma2); % image degradation function
%% Turn RGB to YCbCr, and only SR the Y component
[~, ~, ~, M] = size(LR);
for w = 1:M
LR(:,:,:,w) = rgb2ycbcr(uint8( squeeze(LR(:,:,:,w))));
end
maxiter = 10; % maximum number of iteration
y1(:,:,:) = LR(:,:,1,:);
y2(:,:,:) = LR(:,:,2,:);
y3(:,:,:) = LR(:,:,3,:);
HRitp1 = imresize(y1(:,:,1), spf, 'bicubic'); % bicubic interpolation
HRitp1 = ImWarp(HRitp1, -D(1,1), -D(1,2)); % shift recovering
I1 = Wang_SR(HRitp1, y1, D, Gau, spf, maxiter, reg_term); %Our proposed SR method
HRitp2 = imresize(y2(:,:,1), spf, 'bicubic');
HRitp2 = ImWarp(HRitp2, -D(1,1), -D(1,2));
I2 = HRitp2;
HRitp3 = imresize(y3(:,:,1), spf, 'bicubic');
HRitp3 = ImWarp(HRitp3, -D(1,1), -D(1,2));
I3 = HRitp3;
ImZ(:, :, 1) = I1;
ImZ(:, :, 2) = I2;
ImZ(:, :, 3) = I3;
ImZ = ycbcr2rgb(uint8( ImZ)); % Turn YCbCr to RGB
figure; imshow( uint8( ImZ ) ); title('Wang et al.');
figure; imshow( uint8( IMAGE ) ); title('groundtruth');
%% Evaluation
If = double(ImZ); %output image
Is = double(IMAGE); %reference image
[row,col,~]=size(If);
%RMSE
rmse=0;
for color = 1:3
Ifc = If(:,:,color); Isc = Is(:,:,color);
SSE=sum(sum((Ifc-Isc).^2));
rmsec=sqrt(SSE/(row*col));
rmse = rmse+rmsec/3;
end
rmse
%PSNR
psnr=0;
for color = 1:3
Ifc = If(:,:,color); Isc = Is(:,:,color);
maxIs = max(max(Isc));
minIs = min(min(Isc));
PSNRc = 10*log10((row*col*(maxIs-minIs)^2)/sum(sum((Ifc-Isc).^2)));
psnr = psnr+PSNRc/3;
end
psnr
%SSIM
ssim=0;
for color = 1:3
Ifc = uint8(If(:,:,color)); Isc = uint8(Is(:,:,color));
ssimc = cal_ssim(Ifc, Isc, 0, 0);
ssim = ssim + ssimc/3;
end
ssim
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
- 9.
- 10.
- 11.
- 12.
- 13.
- 14.
- 15.
- 16.
- 17.
- 18.
- 19.
- 20.
- 21.
- 22.
- 23.
- 24.
- 25.
- 26.
- 27.
- 28.
- 29.
- 30.
- 31.
- 32.
- 33.
- 34.
- 35.
- 36.
- 37.
- 38.
- 39.
- 40.
- 41.
- 42.
- 43.
- 44.
- 45.
- 46.
- 47.
- 48.
- 49.
- 50.
- 51.
- 52.
- 53.
- 54.
- 55.
- 56.
- 57.
- 58.
- 59.
- 60.
- 61.
- 62.
- 63.
- 64.
- 65.
- 66.
- 67.
- 68.
- 69.
- 70.
- 71.
- 72.
- 73.
- 74.
- 75.
- 76.
- 77.
- 78.
- 79.
- 80.
- 81.
- 82.
- 83.
- 84.
- 85.
- 86.
- 87.
- 88.
- 89.
- 90.
- 91.
- 92.
- 93.
- 94.
- 95.
- 96.
- 97.
- 98.
- 99.
- 100.
- 101.
- 102.
- 103.
- 104.
- 105.
- 106.
- 107.
- 108.
- 109.
- 110.
3 仿真结果
4 参考文献
[1] Wang Y , Yang J , Xiao C , et al. Fast Convergence Strategy for Multi-image Superresolution via Adaptive Line Search[J]. IEEE Access, 2018:1-1.
5 代码下载