ivan@153
|
1 %% Dictionary Learning for Image Denoising - KSVD vs Recursive Least Squares
|
ivan@153
|
2 %
|
ivan@153
|
3 % This file contains an example of how SMALLbox can be used to test different
|
ivan@153
|
4 % dictionary learning techniques in Image Denoising problem.
|
ivan@153
|
5 % It calls generateImageDenoiseProblem that will let you to choose image,
|
ivan@153
|
6 % add noise and use noisy image to generate training set for dictionary
|
ivan@153
|
7 % learning.
|
ivan@153
|
8 % Two dictionary learning techniques were compared:
|
ivan@153
|
9 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
|
ivan@153
|
10 % Implementation of the K-SVD Algorithm using Batch Orthogonal
|
ivan@153
|
11 % Matching Pursuit", Technical Report - CS, Technion, April 2008.
|
ivan@153
|
12 % - RLS-DLA - Skretting, K.; Engan, K.; , "Recursive Least Squares
|
ivan@153
|
13 % Dictionary Learning Algorithm," Signal Processing, IEEE Transactions on,
|
ivan@153
|
14 % vol.58, no.4, pp.2121-2130, April 2010
|
ivan@153
|
15 %
|
ivan@153
|
16
|
ivan@153
|
17
|
ivan@153
|
18 % Centre for Digital Music, Queen Mary, University of London.
|
ivan@153
|
19 % This file copyright 2011 Ivan Damnjanovic.
|
ivan@153
|
20 %
|
ivan@153
|
21 % This program is free software; you can redistribute it and/or
|
ivan@153
|
22 % modify it under the terms of the GNU General Public License as
|
ivan@153
|
23 % published by the Free Software Foundation; either version 2 of the
|
ivan@153
|
24 % License, or (at your option) any later version. See the file
|
ivan@153
|
25 % COPYING included with this distribution for more information.
|
ivan@153
|
26 %
|
ivan@153
|
27 %%
|
ivan@153
|
28
|
ivan@153
|
29
|
ivan@153
|
30
|
ivan@153
|
31 % If you want to load the image outside of generateImageDenoiseProblem
|
ivan@153
|
32 % function uncomment following lines. This can be useful if you want to
|
ivan@153
|
33 % denoise more then one image for example.
|
ivan@153
|
34 % Here we are loading test_image.mat that contains structure with 5 images : lena,
|
ivan@153
|
35 % barbara,boat, house and peppers.
|
ivan@153
|
36 clear;
|
ivan@153
|
37 TMPpath=pwd;
|
ivan@153
|
38 FS=filesep;
|
luis@186
|
39 [pathstr1, name, ext] = fileparts(which('SMALLboxSetup.m'));
|
ivan@153
|
40 cd([pathstr1,FS,'data',FS,'images']);
|
ivan@153
|
41 load('test_image.mat');
|
ivan@153
|
42 cd(TMPpath);
|
ivan@153
|
43
|
ivan@153
|
44 % Deffining the noise levels that we want to test
|
ivan@153
|
45
|
ivan@153
|
46 noise_level=[10 20 25 50 100];
|
ivan@153
|
47
|
ivan@153
|
48 % Here we loop through different noise levels and images
|
ivan@153
|
49
|
ivan@153
|
50 for noise_ind=2:2
|
ivan@153
|
51 for im_num=2:2
|
ivan@153
|
52
|
ivan@153
|
53 % Defining Image Denoising Problem as Dictionary Learning
|
ivan@153
|
54 % Problem. As an input we set the number of training patches.
|
ivan@153
|
55
|
ivan@153
|
56 SMALL.Problem = generateImageDenoiseProblem(test_image(im_num).i, 40000, '',256, noise_level(noise_ind));
|
ivan@153
|
57 SMALL.Problem.name=int2str(im_num);
|
ivan@153
|
58
|
ivan@153
|
59 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
|
ivan@153
|
60 maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
|
ivan@153
|
61
|
ivan@153
|
62
|
ivan@153
|
63 %%
|
ivan@153
|
64 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
|
ivan@153
|
65 % Boris Mailhe ksvd update implentation omp is the same as with Rubinstein
|
ivan@153
|
66 % implementation
|
ivan@153
|
67
|
ivan@153
|
68
|
ivan@153
|
69 % Initialising solver structure
|
ivan@153
|
70 % Setting solver structure fields (toolbox, name, param, solution,
|
ivan@153
|
71 % reconstructed and time) to zero values
|
ivan@153
|
72
|
ivan@153
|
73 SMALL.solver(1)=SMALL_init_solver;
|
ivan@153
|
74
|
ivan@153
|
75 % Defining the parameters needed for image denoising
|
ivan@153
|
76
|
ivan@153
|
77 SMALL.solver(1).toolbox='ompbox';
|
ivan@153
|
78 SMALL.solver(1).name='omp2';
|
ivan@153
|
79 SMALL.solver(1).param=struct(...
|
ivan@153
|
80 'epsilon',Edata,...
|
ivan@153
|
81 'maxatoms', maxatoms);
|
ivan@153
|
82
|
ivan@153
|
83 % Initialising Dictionary structure
|
ivan@153
|
84 % Setting Dictionary structure fields (toolbox, name, param, D and time)
|
ivan@153
|
85 % to zero values
|
ivan@153
|
86
|
ivan@153
|
87 SMALL.DL(1)=SMALL_init_DL('TwoStepDL', 'KSVD', '', 1);
|
ivan@153
|
88
|
ivan@153
|
89
|
ivan@153
|
90 % Defining the parameters for KSVD
|
ivan@153
|
91 % In this example we are learning 256 atoms in 20 iterations, so that
|
ivan@153
|
92 % every patch in the training set can be represented with target error in
|
ivan@153
|
93 % L2-norm (EData)
|
ivan@153
|
94 % Type help ksvd in MATLAB prompt for more options.
|
ivan@153
|
95
|
ivan@153
|
96
|
ivan@153
|
97 SMALL.DL(1).param=struct(...
|
ivan@153
|
98 'solver', SMALL.solver(1),...
|
ivan@153
|
99 'initdict', SMALL.Problem.initdict,...
|
ivan@153
|
100 'dictsize', SMALL.Problem.p,...
|
ivan@153
|
101 'iternum', 20,...
|
ivan@153
|
102 'show_dict', 1);
|
ivan@153
|
103
|
ivan@153
|
104 % Learn the dictionary
|
ivan@153
|
105
|
ivan@153
|
106 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
|
ivan@153
|
107
|
ivan@153
|
108 % Set SMALL.Problem.A dictionary
|
ivan@153
|
109 % (backward compatiblity with SPARCO: solver structure communicate
|
ivan@153
|
110 % only with Problem structure, ie no direct communication between DL and
|
ivan@153
|
111 % solver structures)
|
ivan@153
|
112
|
ivan@153
|
113 SMALL.Problem.A = SMALL.DL(1).D;
|
ivan@161
|
114 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
|
ivan@153
|
115
|
ivan@153
|
116 % Denoising the image - find the sparse solution in the learned
|
ivan@153
|
117 % dictionary for all patches in the image and the end it uses
|
ivan@153
|
118 % reconstruction function to reconstruct the patches and put them into a
|
ivan@153
|
119 % denoised image
|
ivan@153
|
120
|
ivan@153
|
121 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
|
ivan@153
|
122
|
ivan@153
|
123 %%
|
ivan@153
|
124 % Use MOD Dictionary Learning Algorithm to Learn overcomplete dictionary
|
ivan@153
|
125 % Boris Mailhe MOD update implentation omp is the Ron Rubinstein
|
ivan@153
|
126 % implementation
|
ivan@153
|
127
|
ivan@153
|
128
|
ivan@153
|
129 % Initialising solver structure
|
ivan@153
|
130 % Setting solver structure fields (toolbox, name, param, solution,
|
ivan@153
|
131 % reconstructed and time) to zero values
|
ivan@153
|
132
|
ivan@153
|
133 SMALL.solver(2)=SMALL_init_solver;
|
ivan@153
|
134
|
ivan@153
|
135 % Defining the parameters needed for image denoising
|
ivan@153
|
136
|
ivan@153
|
137 SMALL.solver(2).toolbox='ompbox';
|
ivan@153
|
138 SMALL.solver(2).name='omp2';
|
ivan@153
|
139 SMALL.solver(2).param=struct(...
|
ivan@153
|
140 'epsilon',Edata,...
|
ivan@153
|
141 'maxatoms', maxatoms);
|
ivan@153
|
142
|
ivan@153
|
143 % Initialising Dictionary structure
|
ivan@153
|
144 % Setting Dictionary structure fields (toolbox, name, param, D and time)
|
ivan@153
|
145 % to zero values
|
ivan@153
|
146
|
ivan@153
|
147 SMALL.DL(2)=SMALL_init_DL('TwoStepDL', 'MOD', '', 1);
|
ivan@153
|
148
|
ivan@153
|
149
|
ivan@153
|
150 % Defining the parameters for MOD
|
ivan@153
|
151 % In this example we are learning 256 atoms in 20 iterations, so that
|
ivan@153
|
152 % every patch in the training set can be represented with target error in
|
ivan@153
|
153 % L2-norm (EData)
|
ivan@153
|
154 % Type help ksvd in MATLAB prompt for more options
|
ivan@153
|
155
|
ivan@153
|
156 SMALL.DL(2).param=struct(...
|
ivan@153
|
157 'solver', SMALL.solver(2),...
|
ivan@153
|
158 'initdict', SMALL.Problem.initdict,...
|
ivan@153
|
159 'dictsize', SMALL.Problem.p,...
|
ivan@153
|
160 'iternum', 20,...
|
ivan@153
|
161 'show_dict', 1);
|
ivan@153
|
162
|
ivan@153
|
163 % Learn the dictionary
|
ivan@153
|
164
|
ivan@153
|
165 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2));
|
ivan@153
|
166
|
ivan@153
|
167 % Set SMALL.Problem.A dictionary
|
ivan@153
|
168 % (backward compatiblity with SPARCO: solver structure communicate
|
ivan@153
|
169 % only with Problem structure, ie no direct communication between DL and
|
ivan@153
|
170 % solver structures)
|
ivan@153
|
171
|
ivan@153
|
172 SMALL.Problem.A = SMALL.DL(2).D;
|
ivan@161
|
173 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
|
ivan@153
|
174
|
ivan@153
|
175 % Denoising the image - find the sparse solution in the learned
|
ivan@153
|
176 % dictionary for all patches in the image and the end it uses
|
ivan@153
|
177 % reconstruction function to reconstruct the patches and put them into a
|
ivan@153
|
178 % denoised image
|
ivan@153
|
179
|
ivan@153
|
180 SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
|
ivan@153
|
181 %%
|
ivan@153
|
182 % Use OLS Dictionary Learning Algorithm to Learn overcomplete dictionary
|
ivan@153
|
183 % Boris Mailhe ksvd update implentation omp is the Ron Rubinstein
|
ivan@153
|
184 % implementation
|
ivan@153
|
185
|
ivan@153
|
186
|
ivan@153
|
187 % Initialising solver structure
|
ivan@153
|
188 % Setting solver structure fields (toolbox, name, param, solution,
|
ivan@153
|
189 % reconstructed and time) to zero values
|
ivan@153
|
190
|
ivan@153
|
191 SMALL.solver(3)=SMALL_init_solver;
|
ivan@153
|
192
|
ivan@153
|
193 % Defining the parameters needed for image denoising
|
ivan@153
|
194
|
ivan@153
|
195 SMALL.solver(3).toolbox='ompbox';
|
ivan@153
|
196 SMALL.solver(3).name='omp2';
|
ivan@153
|
197 SMALL.solver(3).param=struct(...
|
ivan@153
|
198 'epsilon',Edata,...
|
ivan@153
|
199 'maxatoms', maxatoms);
|
ivan@153
|
200
|
ivan@153
|
201 % Initialising Dictionary structure
|
ivan@153
|
202 % Setting Dictionary structure fields (toolbox, name, param, D and time)
|
ivan@153
|
203 % to zero values
|
ivan@153
|
204
|
ivan@153
|
205 SMALL.DL(3)=SMALL_init_DL('TwoStepDL', 'ols', '', 1);
|
ivan@153
|
206
|
ivan@153
|
207
|
ivan@153
|
208 % Defining the parameters for KSVD
|
ivan@153
|
209 % In this example we are learning 256 atoms in 20 iterations, so that
|
ivan@153
|
210 % every patch in the training set can be represented with target error in
|
ivan@153
|
211 % L2-norm (EData)
|
ivan@153
|
212 % Type help ksvd in MATLAB prompt for more options.
|
ivan@153
|
213
|
ivan@153
|
214
|
ivan@153
|
215 SMALL.DL(3).param=struct(...
|
ivan@153
|
216 'solver', SMALL.solver(3),...
|
ivan@153
|
217 'initdict', SMALL.Problem.initdict,...
|
ivan@153
|
218 'dictsize', SMALL.Problem.p,...
|
ivan@153
|
219 'iternum', 20,...
|
ivan@153
|
220 'learningRate', 0.1,...
|
ivan@153
|
221 'show_dict', 1);
|
ivan@153
|
222
|
ivan@153
|
223 % Learn the dictionary
|
ivan@153
|
224
|
ivan@153
|
225 SMALL.DL(3) = SMALL_learn(SMALL.Problem, SMALL.DL(3));
|
ivan@153
|
226
|
ivan@153
|
227 % Set SMALL.Problem.A dictionary
|
ivan@153
|
228 % (backward compatiblity with SPARCO: solver structure communicate
|
ivan@153
|
229 % only with Problem structure, ie no direct communication between DL and
|
ivan@153
|
230 % solver structures)
|
ivan@153
|
231
|
ivan@153
|
232 SMALL.Problem.A = SMALL.DL(3).D;
|
ivan@161
|
233 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
|
ivan@153
|
234
|
ivan@153
|
235 % Denoising the image - find the sparse solution in the learned
|
ivan@153
|
236 % dictionary for all patches in the image and the end it uses
|
ivan@153
|
237 % reconstruction function to reconstruct the patches and put them into a
|
ivan@153
|
238 % denoised image
|
ivan@153
|
239
|
ivan@153
|
240 SMALL.solver(3)=SMALL_solve(SMALL.Problem, SMALL.solver(3));
|
ivan@153
|
241 %%
|
ivan@153
|
242 % Use Mailhe Dictionary Learning Algorithm to Learn overcomplete dictionary
|
ivan@153
|
243 % Boris Mailhe ksvd update implentation omp is the Ron Rubinstein
|
ivan@153
|
244 % implementation
|
ivan@153
|
245
|
ivan@153
|
246
|
ivan@153
|
247 % Initialising solver structure
|
ivan@153
|
248 % Setting solver structure fields (toolbox, name, param, solution,
|
ivan@153
|
249 % reconstructed and time) to zero values
|
ivan@153
|
250
|
ivan@153
|
251 SMALL.solver(4)=SMALL_init_solver;
|
ivan@153
|
252
|
ivan@153
|
253 % Defining the parameters needed for image denoising
|
ivan@153
|
254
|
ivan@153
|
255 SMALL.solver(4).toolbox='ompbox';
|
ivan@153
|
256 SMALL.solver(4).name='omp2';
|
ivan@153
|
257 SMALL.solver(4).param=struct(...
|
ivan@153
|
258 'epsilon',Edata,...
|
ivan@153
|
259 'maxatoms', maxatoms);
|
ivan@153
|
260
|
ivan@153
|
261 % Initialising Dictionary structure
|
ivan@153
|
262 % Setting Dictionary structure fields (toolbox, name, param, D and time)
|
ivan@153
|
263 % to zero values
|
ivan@153
|
264
|
bmailhe@212
|
265 SMALL.DL(4)=SMALL_init_DL('TwoStepDL', 'opt', '', 1);
|
ivan@153
|
266
|
ivan@153
|
267
|
ivan@153
|
268 % Defining the parameters for KSVD
|
ivan@153
|
269 % In this example we are learning 256 atoms in 20 iterations, so that
|
ivan@153
|
270 % every patch in the training set can be represented with target error in
|
ivan@153
|
271 % L2-norm (EData)
|
ivan@153
|
272 % Type help ksvd in MATLAB prompt for more options.
|
ivan@153
|
273
|
ivan@153
|
274
|
ivan@153
|
275 SMALL.DL(4).param=struct(...
|
ivan@153
|
276 'solver', SMALL.solver(4),...
|
ivan@153
|
277 'initdict', SMALL.Problem.initdict,...
|
ivan@153
|
278 'dictsize', SMALL.Problem.p,...
|
ivan@153
|
279 'iternum', 20,...
|
ivan@153
|
280 'learningRate', 2,...
|
ivan@153
|
281 'show_dict', 1);
|
ivan@153
|
282
|
ivan@153
|
283 % Learn the dictionary
|
ivan@153
|
284
|
ivan@153
|
285 SMALL.DL(4) = SMALL_learn(SMALL.Problem, SMALL.DL(4));
|
ivan@153
|
286
|
ivan@153
|
287 % Set SMALL.Problem.A dictionary
|
ivan@153
|
288 % (backward compatiblity with SPARCO: solver structure communicate
|
ivan@153
|
289 % only with Problem structure, ie no direct communication between DL and
|
ivan@153
|
290 % solver structures)
|
ivan@153
|
291
|
ivan@153
|
292 SMALL.Problem.A = SMALL.DL(4).D;
|
ivan@161
|
293 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
|
ivan@153
|
294
|
ivan@153
|
295 % Denoising the image - find the sparse solution in the learned
|
ivan@153
|
296 % dictionary for all patches in the image and the end it uses
|
ivan@153
|
297 % reconstruction function to reconstruct the patches and put them into a
|
ivan@153
|
298 % denoised image
|
ivan@153
|
299
|
ivan@153
|
300 SMALL.solver(4)=SMALL_solve(SMALL.Problem, SMALL.solver(4));
|
ivan@153
|
301
|
ivan@153
|
302 %% show results %%
|
ivan@153
|
303
|
ivan@153
|
304 SMALL_ImgDeNoiseResult(SMALL);
|
ivan@153
|
305
|
ivan@153
|
306 %clear SMALL;
|
ivan@153
|
307 end
|
ivan@153
|
308 end
|
ivan@153
|
309
|