Mercurial > hg > smallbox
comparison examples/Image Denoising/SMALL_ImgDenoise_DL_test_KSVDvsSKSVD.m @ 236:5f4e47b78f2b ver_2.0_alpha1
Added example.
author | luisf <luis.figueira@eecs.qmul.ac.uk> |
---|---|
date | Thu, 19 Apr 2012 17:59:08 +0100 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
235:1f5c793c2b18 | 236:5f4e47b78f2b |
---|---|
1 %% Dictionary Learning for Image Denoising - KSVD vs KSVDS vs SPAMS | |
2 % | |
3 % *WARNING!* You should have SPAMS in your search path in order for this | |
4 % script to work.Due to licensing issues SPAMS can not be automatically | |
5 % provided in SMALLbox (http://www.di.ens.fr/willow/SPAMS/downloads.html). | |
6 % | |
7 % This file contains an example of how SMALLbox can be used to test different | |
8 % dictionary learning techniques in Image Denoising problem. | |
9 % It calls generateImageDenoiseProblem that will let you to choose image, | |
10 % add noise and use noisy image to generate training set for dictionary | |
11 % learning. | |
12 % Two dictionary learning techniques were compared: | |
13 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient | |
14 % Implementation of the K-SVD Algorithm using Batch Orthogonal | |
15 % Matching Pursuit", Technical Report - CS, Technion, April 2008. | |
16 % - KSVDS - R. Rubinstein, M. Zibulevsky, and M. Elad, "Learning Sparse | |
17 % Dictionaries for Sparse Signal Approximation", Technical | |
18 % Report - CS, Technion, June 2009. | |
19 % | |
20 | |
21 % | |
22 % Centre for Digital Music, Queen Mary, University of London. | |
23 % This file copyright 2009 Ivan Damnjanovic. | |
24 % | |
25 % This program is free software; you can redistribute it and/or | |
26 % modify it under the terms of the GNU General Public License as | |
27 % published by the Free Software Foundation; either version 2 of the | |
28 % License, or (at your option) any later version. See the file | |
29 % COPYING included with this distribution for more information. | |
30 % | |
31 %% | |
32 | |
33 clear; | |
34 | |
35 % If you want to load the image outside of generateImageDenoiseProblem | |
36 % function uncomment following lines. This can be useful if you want to | |
37 % denoise more then one image for example. | |
38 | |
39 % TMPpath=pwd; | |
40 % FS=filesep; | |
41 % [pathstr1, name, ext] = fileparts(which('SMALLboxSetup.m')); | |
42 % cd([pathstr1,FS,'data',FS,'images']); | |
43 % [filename,pathname] = uigetfile({'*.png;'},'Select a file containin pre-calculated notes'); | |
44 % [pathstr, name, ext] = fileparts(filename); | |
45 % test_image = imread(filename); | |
46 % test_image = double(test_image); | |
47 % cd(TMPpath); | |
48 % SMALL.Problem.name=name; | |
49 | |
50 | |
51 % Defining Image Denoising Problem as Dictionary Learning | |
52 % Problem. As an input we set the number of training patches. | |
53 | |
54 SMALL.Problem = generateImageDenoiseProblem('', 40000); | |
55 | |
56 | |
57 %% | |
58 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary | |
59 | |
60 % Initialising Dictionary structure | |
61 % Setting Dictionary structure fields (toolbox, name, param, D and time) | |
62 % to zero values | |
63 | |
64 SMALL.DL(1)=SMALL_init_DL(); | |
65 | |
66 % Defining the parameters needed for dictionary learning | |
67 | |
68 SMALL.DL(1).toolbox = 'KSVD'; | |
69 SMALL.DL(1).name = 'ksvd'; | |
70 | |
71 % Defining the parameters for KSVD | |
72 % In this example we are learning 256 atoms in 20 iterations, so that | |
73 % every patch in the training set can be represented with target error in | |
74 % L2-norm (EData) | |
75 % Type help ksvd in MATLAB prompt for more options. | |
76 | |
77 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain; | |
78 maxatoms = floor(prod(SMALL.Problem.blocksize)/2); | |
79 | |
80 SMALL.DL(1).param=struct(... | |
81 'Edata', Edata,... | |
82 'initdict', SMALL.Problem.initdict,... | |
83 'dictsize', SMALL.Problem.p,... | |
84 'iternum', 20,... | |
85 'memusage', 'high'); | |
86 | |
87 % Learn the dictionary | |
88 | |
89 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1)); | |
90 | |
91 % Set SMALL.Problem.A dictionary | |
92 % (backward compatiblity with SPARCO: solver structure communicate | |
93 % only with Problem structure, ie no direct communication between DL and | |
94 % solver structures) | |
95 | |
96 SMALL.Problem.A = SMALL.DL(1).D; | |
97 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem); | |
98 | |
99 %% | |
100 % Initialising solver structure | |
101 % Setting solver structure fields (toolbox, name, param, solution, | |
102 % reconstructed and time) to zero values | |
103 | |
104 SMALL.solver(1)=SMALL_init_solver; | |
105 | |
106 % Defining the parameters needed for image denoising | |
107 | |
108 SMALL.solver(1).toolbox='ompbox'; | |
109 SMALL.solver(1).name='omp2'; | |
110 SMALL.solver(1).param=struct(... | |
111 'epsilon',Edata,... | |
112 'maxatoms', maxatoms); | |
113 | |
114 % Denoising the image - find the sparse solution in the learned | |
115 % dictionary for all patches in the image and the end it uses | |
116 % reconstruction function to reconstruct the patches and put them into a | |
117 % denoised image | |
118 | |
119 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1)); | |
120 | |
121 % Show PSNR after reconstruction | |
122 | |
123 SMALL.solver(1).reconstructed.psnr | |
124 | |
125 %% | |
126 % Use KSVDS Dictionary Learning Algorithm to denoise image | |
127 | |
128 % Initialising solver structure | |
129 % Setting solver structure fields (toolbox, name, param, solution, | |
130 % reconstructed and time) to zero values | |
131 | |
132 SMALL.DL(2)=SMALL_init_DL(); | |
133 | |
134 % Defining the parameters needed for dictionary learning | |
135 | |
136 SMALL.DL(2).toolbox = 'KSVDS'; | |
137 SMALL.DL(2).name = 'ksvds'; | |
138 | |
139 % Defining the parameters for KSVDS | |
140 % In this example we are learning 256 atoms in 20 iterations, so that | |
141 % every patch in the training set can be represented with target error in | |
142 % L2-norm (EDataS). We also impose "double sparsity" - dictionary itself | |
143 % has to be sparse in the given base dictionary (Tdict - number of | |
144 % nonzero elements per atom). | |
145 % Type help ksvds in MATLAB prompt for more options. | |
146 | |
147 EdataS=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain; | |
148 SMALL.DL(2).param=struct(... | |
149 'Edata', EdataS, ... | |
150 'Tdict', 6,... | |
151 'stepsize', 1,... | |
152 'dictsize', SMALL.Problem.p,... | |
153 'iternum', 20,... | |
154 'memusage', 'high'); | |
155 SMALL.DL(2).param.initA = speye(SMALL.Problem.p); | |
156 SMALL.DL(2).param.basedict{1} = odctdict(8,16); | |
157 SMALL.DL(2).param.basedict{2} = odctdict(8,16); | |
158 | |
159 % Learn the dictionary | |
160 | |
161 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2)); | |
162 | |
163 % Set SMALL.Problem.A dictionary and SMALL.Problem.basedictionary | |
164 % (backward compatiblity with SPARCO: solver structure communicate | |
165 % only with Problem structure, ie no direct communication between DL and | |
166 % solver structures) | |
167 | |
168 SMALL.Problem.A = SMALL.DL(2).D; | |
169 SMALL.Problem.basedict{1} = SMALL.DL(2).param.basedict{1}; | |
170 SMALL.Problem.basedict{2} = SMALL.DL(2).param.basedict{2}; | |
171 | |
172 % Setting up reconstruction function | |
173 | |
174 SparseDict=1; | |
175 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem, SparseDict); | |
176 | |
177 % Initialising solver structure | |
178 % Setting solver structure fields (toolbox, name, param, solution, | |
179 % reconstructed and time) to zero values | |
180 | |
181 SMALL.solver(2)=SMALL_init_solver; | |
182 | |
183 % Defining the parameters needed for image denoising | |
184 | |
185 SMALL.solver(2).toolbox='ompsbox'; | |
186 SMALL.solver(2).name='omps2'; | |
187 SMALL.solver(2).param=struct(... | |
188 'epsilon',Edata,... | |
189 'maxatoms', maxatoms); | |
190 | |
191 % Denoising the image - find the sparse solution in the learned | |
192 % dictionary for all patches in the image and the end it uses | |
193 % reconstruction function to reconstruct the patches and put them into a | |
194 % denoised image | |
195 | |
196 SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2)); | |
197 | |
198 | |
199 %% | |
200 % Plot results and save midi files | |
201 | |
202 % show results % | |
203 | |
204 SMALL_ImgDeNoiseResult(SMALL); |