-
Notifications
You must be signed in to change notification settings - Fork 0
/
Demo_genearal_degradation_SRMD.m
221 lines (166 loc) · 8.03 KB
/
Demo_genearal_degradation_SRMD.m
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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
%==========================================================================
% This is the testing code of SRMD for the <general degradation> of SISR.
% For general degradation, the basic setting is:
% 1. there are tree types of kernels, including isotropic Gaussian,
% anisotropic Gaussian, and estimated kernel k_b for isotropic
% Gaussian k_d under direct downsampler (x2 and x3 only).
% 2. the noise level range is [0, 75].
% 3. the downsampler is fixed to bicubic downsampler.
% For direct downsampler, you can either train a new model with
% direct downsamper or use the estimated kernel k_b under direct
% downsampler. The former is preferred.
% 4. there are three models, "SRMDx2.mat" for scale factor 2, "SRMDx3.mat"
% for scale factor 3, and "SRMDx4.mat" for scale factor 4.
%==========================================================================
% The basic idea of SRMD is to learn a CNN to infer the MAP of general SISR, i.e.,
% solve x^ = arg min_x 1/(2 sigma^2) ||(kx)\downarrow_s - y||^2 + lamda \Phi(x)
% via x^ = CNN(y,k,sigma;\Theta) or HR^ = CNN(LR,kernel,noiselevel;\Theta).
%
% There involves two important factors, i.e., blur kernel (k; kernel) and noise
% level (sigma; nlevel).
%
% For more information, please refer to the following paper.
% @article{zhang2017learningsrmd,
% title={Learning a Single Convolutional Super-Resolution Network for Multiple Degradations},
% author={Kai, Zhang and Wangmeng, Zuo and Lei, Zhang},
% year={2017},
% }
%
% If you have any question, please feel free to contact with <Kai Zhang ([email protected])>.
%
% This code is for research purpose only.
%
% by Kai Zhang (Nov, 2017)
%==========================================================================
% clear; clc;
format compact;
addpath('utilities');
imageSets = {'Set5','Set14','BSD100','Urban100'}; % testing dataset
%% select testing dataset, use GPU or not, ...
setTest = imageSets([1]); %
showResult = 1; % 1, show ground-truth, bicubicly interpolated LR image, and restored HR images by SRMD; 2, save restored images
pauseTime = 1;
useGPU = 1; % 1 or 0, true or false
method = 'SRMD';
folderTest = 'testsets';
folderResult = 'results';
if ~exist(folderResult,'file')
mkdir(folderResult);
end
%% scale factor (2, 3, 4)
sf = 2; %{2, 3, 4}
%% load model with scale factor sf
folderModel = 'models';
load(fullfile(folderModel,['SRMDx',int2str(sf),'.mat']));
%net.layers = net.layers(1:end-1);
net = vl_simplenn_tidy(net);
if useGPU
net = vl_simplenn_move(net, 'gpu') ;
end
%% degradation parameter (noise level and kernel) setting
%############################# noise level ################################
% noise level, from a range of [0, 75]
nlevel = 15; % [0, 75]
kerneltype = 1; % {1, 2, 3}
%############################### kernel ###################################
% there are tree types of kernels, including isotropic Gaussian,
% anisotropic Gaussian, and estimated kernel k_b for isotropic Gaussian k_d
% under direct downsampler (x2 and x3 only).
if kerneltype == 1
% type 1, isotropic Gaussian---although it is a special case of anisotropic Gaussian.
kernelwidth = 2.6; % from a range of [0.2, 2] for sf = 2, [0.2, 3] for sf = 3, and [0.2, 4] for sf = 4.
kernel = fspecial('gaussian',15, kernelwidth); % Note: the kernel size is fixed to 15X15.
tag = ['_',method,'_x',num2str(sf),'_itrG_',int2str(kernelwidth*10),'_nlevel_',int2str(nlevel)];
elseif kerneltype == 2
% type 2, anisotropic Gaussian
nk = randi(size(net.meta.AtrpGaussianKernel,4)); % randomly select one
kernel = net.meta.AtrpGaussianKernel(:,:,:,nk);
tag = ['_',method,'_x',num2str(sf),'_atrG_',int2str(nk),'_nlevel_',int2str(nlevel)];
elseif kerneltype == 3 && ( sf==2 || sf==3 )
% type 3, estimated kernel k_b (x2 and x3 only)
nk = randi(size(net.meta.directKernel,4)); % randomly select one
kernel = net.meta.directKernel(:,:,:,nk);
tag = ['_',method,'_x',num2str(sf),'_dirG_',int2str(nk),'_nlevel_',int2str(nlevel)];
end
%##########################################################################
surf(kernel) % show kernel
view(45,55);
title('Assumed kernel');
xlim([1 15]);
ylim([1 15]);
pause(2)
close;
%% for degradation maps
global degpar;
degpar = single([net.meta.P*kernel(:); nlevel(:)/255]);
for n_set = 1 : numel(setTest)
%% search images
setTestCur = cell2mat(setTest(n_set));
disp('--------------------------------------------');
disp([' ----',setTestCur,'-----Super-Resolution-----']);
disp('--------------------------------------------');
folderTestCur = fullfile(folderTest,setTestCur);
ext = {'*.jpg','*.png','*.bmp'};
filepaths = [];
for i = 1 : length(ext)
filepaths = cat(1,filepaths,dir(fullfile(folderTestCur, ext{i})));
end
%% prepare results
eval(['PSNR_',setTestCur,'_x',num2str(sf),' = zeros(length(filepaths),1);']);
eval(['SSIM_',setTestCur,'_x',num2str(sf),' = zeros(length(filepaths),1);']);
folderResultCur = fullfile(folderResult, [setTestCur,tag]);
if ~exist(folderResultCur,'file')
mkdir(folderResultCur);
end
%% perform SISR
for i = 1 : length(filepaths)
HR = imread(fullfile(folderTestCur,filepaths(i).name));
C = size(HR,3);
if C == 1
HR = cat(3,HR,HR,HR);
end
[~,imageName,ext] = fileparts(filepaths(i).name);
HR = modcrop(HR, sf);
label_RGB = HR;
blury_HR = imfilter(im2double(HR),double(kernel),'replicate'); % blur
LR = imresize(blury_HR,1/sf,'bicubic'); % bicubic downsampling
randn('seed',0);
LR_noisy = LR + nlevel/255.*randn(size(LR)); % add random noise (AWGN)
input = single(LR_noisy);
%tic
if useGPU
input = gpuArray(input);
end
res = vl_srmd(net, input,[],[],'conserveMemory',true,'mode','test','cudnn',true);
%res = vl_srmd_concise(net, input); % a concise version of "vl_srmd".
%res = vl_srmd_matlab(net, input); % When use this, you should also set "useGPU = 0;" and comment "net = vl_simplenn_tidy(net);"
output_RGB = gather(res(end).x);
%toc;
if C == 1
label = mean(im2double(HR),3);
output = mean(output_RGB,3);
else
label = rgb2ycbcr(im2double(HR));
output = rgb2ycbcr(double(output_RGB));
label = label(:,:,1);
output = output(:,:,1);
end
%% calculate PSNR and SSIM
[PSNR_Cur,SSIM_Cur] = Cal_PSNRSSIM(label*255,output*255,sf,sf); %%% single
disp([setTestCur,' ',int2str(i),' ',num2str(PSNR_Cur,'%2.2f'),'dB',' ',filepaths(i).name]);
eval(['PSNR_',setTestCur,'_x',num2str(sf),'(',num2str(i),') = PSNR_Cur;']);
eval(['SSIM_',setTestCur,'_x',num2str(sf),'(',num2str(i),') = SSIM_Cur;']);
if showResult
imshow(cat(2,label_RGB,imresize(im2uint8(LR_noisy),sf),im2uint8(output_RGB)));
drawnow;
title(['SISR ',filepaths(i).name,' ',num2str(PSNR_Cur,'%2.2f'),'dB'],'FontSize',12)
pause(pauseTime)
imwrite(output_RGB,fullfile(folderResultCur,[imageName,'_x',int2str(sf),'_',int2str(PSNR_Cur*100),'.png']));% save results
end
end
disp(['Average PSNR is ',num2str(mean(eval(['PSNR_',setTestCur,'_x',num2str(sf)])),'%2.2f'),'dB']);
disp(['Average SSIM is ',num2str(mean(eval(['SSIM_',setTestCur,'_x',num2str(sf)])),'%2.4f')]);
%% save PSNR and SSIM results
save(fullfile(folderResultCur,['PSNR_',setTestCur,'_x',num2str(sf),'.mat']),['PSNR_',setTestCur,'_x',num2str(sf)]);
save(fullfile(folderResultCur,['SSIM_',setTestCur,'_x',num2str(sf),'.mat']),['SSIM_',setTestCur,'_x',num2str(sf)]);
end