comparison examples/Image Denoising/SMALL_ImgDenoise_DL_test_KSVDvsRLSDLAvsTwoStepMOD.m @ 152:485747bf39e0 ivand_dev

Two step dictonary learning - Integration of the code for dictionary update and dictionary decorrelation from Boris Mailhe
author Ivan Damnjanovic lnx <ivan.damnjanovic@eecs.qmul.ac.uk>
date Thu, 28 Jul 2011 15:49:32 +0100
parents
children af307f247ac7
comparison
equal deleted inserted replaced
149:fec205ec6ef6 152:485747bf39e0
1 %% Dictionary Learning for Image Denoising - KSVD vs Recursive Least Squares
2 %
3 % This file contains an example of how SMALLbox can be used to test different
4 % dictionary learning techniques in Image Denoising problem.
5 % It calls generateImageDenoiseProblem that will let you to choose image,
6 % add noise and use noisy image to generate training set for dictionary
7 % learning.
8 % Two dictionary learning techniques were compared:
9 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
10 % Implementation of the K-SVD Algorithm using Batch Orthogonal
11 % Matching Pursuit", Technical Report - CS, Technion, April 2008.
12 % - RLS-DLA - Skretting, K.; Engan, K.; , "Recursive Least Squares
13 % Dictionary Learning Algorithm," Signal Processing, IEEE Transactions on,
14 % vol.58, no.4, pp.2121-2130, April 2010
15 %
16
17
18 % Centre for Digital Music, Queen Mary, University of London.
19 % This file copyright 2011 Ivan Damnjanovic.
20 %
21 % This program is free software; you can redistribute it and/or
22 % modify it under the terms of the GNU General Public License as
23 % published by the Free Software Foundation; either version 2 of the
24 % License, or (at your option) any later version. See the file
25 % COPYING included with this distribution for more information.
26 %
27 %%
28
29
30
31 % If you want to load the image outside of generateImageDenoiseProblem
32 % function uncomment following lines. This can be useful if you want to
33 % denoise more then one image for example.
34 % Here we are loading test_image.mat that contains structure with 5 images : lena,
35 % barbara,boat, house and peppers.
36 clear;
37 TMPpath=pwd;
38 FS=filesep;
39 [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
40 cd([pathstr1,FS,'data',FS,'images']);
41 load('test_image.mat');
42 cd(TMPpath);
43
44 % Deffining the noise levels that we want to test
45
46 noise_level=[10 20 25 50 100];
47
48 % Here we loop through different noise levels and images
49
50 for noise_ind=2:2
51 for im_num=1:1
52
53 % Defining Image Denoising Problem as Dictionary Learning
54 % Problem. As an input we set the number of training patches.
55
56 SMALL.Problem = generateImageDenoiseProblem(test_image(im_num).i, 40000, '',256, noise_level(noise_ind));
57 SMALL.Problem.name=int2str(im_num);
58
59 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
60 maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
61
62 % results structure is to store all results
63
64 results(noise_ind,im_num).noisy_psnr=SMALL.Problem.noisy_psnr;
65
66 %%
67 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
68
69 % Initialising Dictionary structure
70 % Setting Dictionary structure fields (toolbox, name, param, D and time)
71 % to zero values
72
73 SMALL.DL(1)=SMALL_init_DL();
74
75 % Defining the parameters needed for dictionary learning
76
77 SMALL.DL(1).toolbox = 'KSVD';
78 SMALL.DL(1).name = 'ksvd';
79
80 % Defining the parameters for KSVD
81 % In this example we are learning 256 atoms in 20 iterations, so that
82 % every patch in the training set can be represented with target error in
83 % L2-norm (Edata)
84 % Type help ksvd in MATLAB prompt for more options.
85
86
87 SMALL.DL(1).param=struct(...
88 'Edata', Edata,...
89 'initdict', SMALL.Problem.initdict,...
90 'dictsize', SMALL.Problem.p,...
91 'exact', 1, ...
92 'iternum', 20,...
93 'memusage', 'high');
94
95 % Learn the dictionary
96
97 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
98
99 % Set SMALL.Problem.A dictionary
100 % (backward compatiblity with SPARCO: solver structure communicate
101 % only with Problem structure, ie no direct communication between DL and
102 % solver structures)
103
104 SMALL.Problem.A = SMALL.DL(1).D;
105 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
106
107 %%
108 % Initialising solver structure
109 % Setting solver structure fields (toolbox, name, param, solution,
110 % reconstructed and time) to zero values
111
112 SMALL.solver(1)=SMALL_init_solver;
113
114 % Defining the parameters needed for image denoising
115
116 SMALL.solver(1).toolbox='ompbox';
117 SMALL.solver(1).name='omp2';
118 SMALL.solver(1).param=struct(...
119 'epsilon',Edata,...
120 'maxatoms', maxatoms);
121
122 % Denoising the image - find the sparse solution in the learned
123 % dictionary for all patches in the image and the end it uses
124 % reconstruction function to reconstruct the patches and put them into a
125 % denoised image
126
127 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
128
129 % Show PSNR after reconstruction
130
131 SMALL.solver(1).reconstructed.psnr
132
133 %%
134 % For comparison purposes we will denoise image with overcomplete DCT
135 % here
136 % Set SMALL.Problem.A dictionary to be oDCT (i.e. Problem.initdict -
137 % since initial dictionaruy is already set to be oDCT when generating the
138 % denoising problem
139
140
141 % Initialising solver structure
142 % Setting solver structure fields (toolbox, name, param, solution,
143 % reconstructed and time) to zero values
144
145 SMALL.solver(2)=SMALL_init_solver;
146
147 % Defining the parameters needed for image denoising
148
149 SMALL.solver(2).toolbox='ompbox';
150 SMALL.solver(2).name='omp2';
151 SMALL.solver(2).param=struct(...
152 'epsilon',Edata,...
153 'maxatoms', maxatoms);
154
155 % Initialising Dictionary structure
156 % Setting Dictionary structure fields (toolbox, name, param, D and time)
157 % to zero values
158
159 SMALL.DL(2)=SMALL_init_DL('TwoStepDL', 'MOD', '', 1);
160
161
162 % Defining the parameters for MOD
163 % In this example we are learning 256 atoms in 20 iterations, so that
164 % every patch in the training set can be represented with target error in
165 % L2-norm (EData)
166 % Type help ksvd in MATLAB prompt for more options.
167
168
169 SMALL.DL(2).param=struct(...
170 'solver', SMALL.solver(2),...
171 'initdict', SMALL.Problem.initdict,...
172 'dictsize', SMALL.Problem.p,...
173 'iternum', 40,...
174 'mu', 0.7,...
175 'show_dict', 1);
176
177 % Learn the dictionary
178
179 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2));
180
181 % Set SMALL.Problem.A dictionary
182 % (backward compatiblity with SPARCO: solver structure communicate
183 % only with Problem structure, ie no direct communication between DL and
184 % solver structures)
185
186 SMALL.Problem.A = SMALL.DL(2).D;
187 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
188
189 % Denoising the image - find the sparse solution in the learned
190 % dictionary for all patches in the image and the end it uses
191 % reconstruction function to reconstruct the patches and put them into a
192 % denoised image
193
194 SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
195
196 %%
197 % In the b1 field all patches from the image are stored. For RLS-DLA we
198 % will first exclude all the patches that have l2 norm smaller then
199 % threshold and then take min(40000, number_of_remaining_patches) in
200 % ascending order as our training set (SMALL.Problem.b)
201
202 X=SMALL.Problem.b1;
203 X_norm=sqrt(sum(X.^2, 1));
204 [X_norm_sort, p]=sort(X_norm);
205 p1=p(X_norm_sort>Edata);
206 if size(p1,2)>40000
207 p2 = randperm(size(p1,2));
208 p2=sort(p2(1:40000));
209 size(p2,2)
210 SMALL.Problem.b=X(:,p1(p2));
211 else
212 size(p1,2)
213 SMALL.Problem.b=X(:,p1);
214
215 end
216
217 % Forgetting factor for RLS-DLA algorithm, in this case we are using
218 % fixed value
219
220 lambda=0.9998
221
222 % Use Recursive Least Squares
223 % to Learn overcomplete dictionary
224
225 % Initialising Dictionary structure
226 % Setting Dictionary structure fields (toolbox, name, param, D and time)
227 % to zero values
228
229 SMALL.DL(3)=SMALL_init_DL();
230
231 % Defining fields needed for dictionary learning
232
233 SMALL.DL(3).toolbox = 'SMALL';
234 SMALL.DL(3).name = 'SMALL_rlsdla';
235 SMALL.DL(3).param=struct(...
236 'Edata', Edata,...
237 'initdict', SMALL.Problem.initdict,...
238 'dictsize', SMALL.Problem.p,...
239 'forgettingMode', 'FIX',...
240 'forgettingFactor', lambda,...
241 'show_dict', 1000);
242
243
244 SMALL.DL(3) = SMALL_learn(SMALL.Problem, SMALL.DL(3));
245
246 % Initialising solver structure
247 % Setting solver structure fields (toolbox, name, param, solution,
248 % reconstructed and time) to zero values
249
250 SMALL.Problem.A = SMALL.DL(3).D;
251 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
252
253 SMALL.solver(3)=SMALL_init_solver;
254
255 % Defining the parameters needed for image denoising
256
257 SMALL.solver(3).toolbox='ompbox';
258 SMALL.solver(3).name='omp2';
259 SMALL.solver(3).param=struct(...
260 'epsilon',Edata,...
261 'maxatoms', maxatoms);
262
263
264 SMALL.solver(3)=SMALL_solve(SMALL.Problem, SMALL.solver(3));
265
266 SMALL.solver(3).reconstructed.psnr
267
268
269 % show results %
270
271 SMALL_ImgDeNoiseResult(SMALL);
272
273 results(noise_ind,im_num).psnr.ksvd=SMALL.solver(1).reconstructed.psnr;
274 results(noise_ind,im_num).psnr.odct=SMALL.solver(2).reconstructed.psnr;
275 results(noise_ind,im_num).psnr.rlsdla=SMALL.solver(3).reconstructed.psnr;
276 results(noise_ind,im_num).vmrse.ksvd=SMALL.solver(1).reconstructed.vmrse;
277 results(noise_ind,im_num).vmrse.odct=SMALL.solver(2).reconstructed.vmrse;
278 results(noise_ind,im_num).vmrse.rlsdla=SMALL.solver(3).reconstructed.vmrse;
279 results(noise_ind,im_num).ssim.ksvd=SMALL.solver(1).reconstructed.ssim;
280 results(noise_ind,im_num).ssim.odct=SMALL.solver(2).reconstructed.ssim;
281 results(noise_ind,im_num).ssim.rlsdla=SMALL.solver(3).reconstructed.ssim;
282
283 results(noise_ind,im_num).time.ksvd=SMALL.solver(1).time+SMALL.DL(1).time;
284 results(noise_ind,im_num).time.rlsdla.time=SMALL.solver(3).time+SMALL.DL(3).time;
285 clear SMALL;
286 end
287 end
288 % save results.mat results