annotate examples/Image Denoising/SMALL_ImgDenoise_DL_test_KSVDvsSPAMS.m @ 217:8b3c71bb44eb luisf_dev

Removed "clear all" from example scripts (subs by "clear" instead)
author luisf <luis.figueira@eecs.qmul.ac.uk>
date Thu, 22 Mar 2012 14:41:04 +0000
parents 9c418bea7f6a
children
rev   line source
ivan@128 1 %% Dictionary Learning for Image Denoising - KSVD vs KSVDS vs SPAMS
ivan@128 2 %
ivan@128 3 % *WARNING!* You should have SPAMS in your search path in order for this
ivan@128 4 % script to work.Due to licensing issues SPAMS can not be automatically
ivan@128 5 % provided in SMALLbox (http://www.di.ens.fr/willow/SPAMS/downloads.html).
ivan@128 6 %
idamnjanovic@6 7 % This file contains an example of how SMALLbox can be used to test different
idamnjanovic@6 8 % dictionary learning techniques in Image Denoising problem.
idamnjanovic@6 9 % It calls generateImageDenoiseProblem that will let you to choose image,
idamnjanovic@6 10 % add noise and use noisy image to generate training set for dictionary
idamnjanovic@6 11 % learning.
idamnjanovic@6 12 % Three dictionary learning techniques were compared:
idamnjanovic@6 13 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
idamnjanovic@6 14 % Implementation of the K-SVD Algorithm using Batch Orthogonal
idamnjanovic@6 15 % Matching Pursuit", Technical Report - CS, Technion, April 2008.
idamnjanovic@6 16 % - KSVDS - R. Rubinstein, M. Zibulevsky, and M. Elad, "Learning Sparse
idamnjanovic@6 17 % Dictionaries for Sparse Signal Approximation", Technical
idamnjanovic@6 18 % Report - CS, Technion, June 2009.
idamnjanovic@6 19 % - SPAMS - J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online
idamnjanovic@6 20 % Dictionary Learning for Sparse Coding. International
idamnjanovic@6 21 % Conference on Machine Learning,Montreal, Canada, 2009
idamnjanovic@6 22 %
ivan@107 23
ivan@107 24 %
ivan@107 25 % Centre for Digital Music, Queen Mary, University of London.
ivan@107 26 % This file copyright 2009 Ivan Damnjanovic.
ivan@107 27 %
ivan@107 28 % This program is free software; you can redistribute it and/or
ivan@107 29 % modify it under the terms of the GNU General Public License as
ivan@107 30 % published by the Free Software Foundation; either version 2 of the
ivan@107 31 % License, or (at your option) any later version. See the file
ivan@107 32 % COPYING included with this distribution for more information.
idamnjanovic@6 33 %
idamnjanovic@6 34 %%
idamnjanovic@6 35
idamnjanovic@6 36 clear;
idamnjanovic@6 37
idamnjanovic@6 38 % If you want to load the image outside of generateImageDenoiseProblem
idamnjanovic@6 39 % function uncomment following lines. This can be useful if you want to
idamnjanovic@6 40 % denoise more then one image for example.
idamnjanovic@6 41
idamnjanovic@6 42 % TMPpath=pwd;
idamnjanovic@6 43 % FS=filesep;
luis@186 44 % [pathstr1, name, ext] = fileparts(which('SMALLboxSetup.m'));
idamnjanovic@6 45 % cd([pathstr1,FS,'data',FS,'images']);
idamnjanovic@6 46 % [filename,pathname] = uigetfile({'*.png;'},'Select a file containin pre-calculated notes');
luis@186 47 % [pathstr, name, ext] = fileparts(filename);
idamnjanovic@6 48 % test_image = imread(filename);
idamnjanovic@6 49 % test_image = double(test_image);
idamnjanovic@6 50 % cd(TMPpath);
idamnjanovic@6 51 % SMALL.Problem.name=name;
idamnjanovic@6 52
idamnjanovic@6 53
idamnjanovic@6 54 % Defining Image Denoising Problem as Dictionary Learning
idamnjanovic@6 55 % Problem. As an input we set the number of training patches.
idamnjanovic@6 56
idamnjanovic@6 57 SMALL.Problem = generateImageDenoiseProblem('', 40000);
idamnjanovic@6 58
idamnjanovic@6 59
idamnjanovic@6 60 %%
idamnjanovic@6 61 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
idamnjanovic@6 62
idamnjanovic@6 63 % Initialising Dictionary structure
idamnjanovic@6 64 % Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@6 65 % to zero values
idamnjanovic@6 66
idamnjanovic@6 67 SMALL.DL(1)=SMALL_init_DL();
idamnjanovic@6 68
idamnjanovic@6 69 % Defining the parameters needed for dictionary learning
idamnjanovic@6 70
idamnjanovic@6 71 SMALL.DL(1).toolbox = 'KSVD';
idamnjanovic@6 72 SMALL.DL(1).name = 'ksvd';
idamnjanovic@6 73
idamnjanovic@6 74 % Defining the parameters for KSVD
idamnjanovic@6 75 % In this example we are learning 256 atoms in 20 iterations, so that
idamnjanovic@6 76 % every patch in the training set can be represented with target error in
idamnjanovic@6 77 % L2-norm (EData)
idamnjanovic@6 78 % Type help ksvd in MATLAB prompt for more options.
idamnjanovic@6 79
idamnjanovic@6 80 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
ivan@107 81 maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
ivan@107 82
idamnjanovic@6 83 SMALL.DL(1).param=struct(...
idamnjanovic@6 84 'Edata', Edata,...
idamnjanovic@6 85 'initdict', SMALL.Problem.initdict,...
idamnjanovic@6 86 'dictsize', SMALL.Problem.p,...
idamnjanovic@6 87 'iternum', 20,...
idamnjanovic@6 88 'memusage', 'high');
idamnjanovic@6 89
idamnjanovic@6 90 % Learn the dictionary
idamnjanovic@6 91
idamnjanovic@6 92 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
idamnjanovic@6 93
idamnjanovic@6 94 % Set SMALL.Problem.A dictionary
idamnjanovic@6 95 % (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@6 96 % only with Problem structure, ie no direct communication between DL and
idamnjanovic@6 97 % solver structures)
idamnjanovic@6 98
idamnjanovic@6 99 SMALL.Problem.A = SMALL.DL(1).D;
ivan@161 100 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
idamnjanovic@6 101
idamnjanovic@6 102 %%
idamnjanovic@6 103 % Initialising solver structure
idamnjanovic@6 104 % Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@6 105 % reconstructed and time) to zero values
idamnjanovic@6 106
idamnjanovic@6 107 SMALL.solver(1)=SMALL_init_solver;
idamnjanovic@6 108
idamnjanovic@6 109 % Defining the parameters needed for image denoising
idamnjanovic@6 110
idamnjanovic@6 111 SMALL.solver(1).toolbox='ompbox';
ivan@107 112 SMALL.solver(1).name='omp2';
ivan@107 113 SMALL.solver(1).param=struct(...
ivan@107 114 'epsilon',Edata,...
ivan@107 115 'maxatoms', maxatoms);
idamnjanovic@6 116
ivan@107 117 % Denoising the image - find the sparse solution in the learned
ivan@107 118 % dictionary for all patches in the image and the end it uses
ivan@107 119 % reconstruction function to reconstruct the patches and put them into a
ivan@107 120 % denoised image
idamnjanovic@6 121
ivan@107 122 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
ivan@107 123
ivan@107 124 % Show PSNR after reconstruction
ivan@107 125
ivan@107 126 SMALL.solver(1).reconstructed.psnr
idamnjanovic@6 127
idamnjanovic@6 128 %%
idamnjanovic@6 129 % Use KSVDS Dictionary Learning Algorithm to denoise image
idamnjanovic@6 130
idamnjanovic@6 131 % Initialising solver structure
idamnjanovic@6 132 % Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@6 133 % reconstructed and time) to zero values
idamnjanovic@6 134
idamnjanovic@6 135 SMALL.DL(2)=SMALL_init_DL();
idamnjanovic@6 136
idamnjanovic@6 137 % Defining the parameters needed for dictionary learning
idamnjanovic@6 138
idamnjanovic@6 139 SMALL.DL(2).toolbox = 'KSVDS';
idamnjanovic@6 140 SMALL.DL(2).name = 'ksvds';
idamnjanovic@6 141
idamnjanovic@6 142 % Defining the parameters for KSVDS
idamnjanovic@6 143 % In this example we are learning 256 atoms in 20 iterations, so that
idamnjanovic@6 144 % every patch in the training set can be represented with target error in
idamnjanovic@6 145 % L2-norm (EDataS). We also impose "double sparsity" - dictionary itself
idamnjanovic@6 146 % has to be sparse in the given base dictionary (Tdict - number of
idamnjanovic@6 147 % nonzero elements per atom).
idamnjanovic@6 148 % Type help ksvds in MATLAB prompt for more options.
idamnjanovic@6 149
idamnjanovic@6 150 EdataS=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
idamnjanovic@6 151 SMALL.DL(2).param=struct(...
idamnjanovic@6 152 'Edata', EdataS, ...
idamnjanovic@6 153 'Tdict', 6,...
idamnjanovic@6 154 'stepsize', 1,...
idamnjanovic@6 155 'dictsize', SMALL.Problem.p,...
idamnjanovic@6 156 'iternum', 20,...
idamnjanovic@6 157 'memusage', 'high');
idamnjanovic@6 158 SMALL.DL(2).param.initA = speye(SMALL.Problem.p);
idamnjanovic@6 159 SMALL.DL(2).param.basedict{1} = odctdict(8,16);
idamnjanovic@6 160 SMALL.DL(2).param.basedict{2} = odctdict(8,16);
idamnjanovic@6 161
idamnjanovic@6 162 % Learn the dictionary
idamnjanovic@6 163
idamnjanovic@6 164 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2));
idamnjanovic@6 165
idamnjanovic@6 166 % Set SMALL.Problem.A dictionary and SMALL.Problem.basedictionary
idamnjanovic@6 167 % (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@6 168 % only with Problem structure, ie no direct communication between DL and
idamnjanovic@6 169 % solver structures)
idamnjanovic@6 170
idamnjanovic@6 171 SMALL.Problem.A = SMALL.DL(2).D;
idamnjanovic@6 172 SMALL.Problem.basedict{1} = SMALL.DL(2).param.basedict{1};
idamnjanovic@6 173 SMALL.Problem.basedict{2} = SMALL.DL(2).param.basedict{2};
idamnjanovic@6 174
ivan@107 175 % Setting up reconstruction function
ivan@107 176
ivan@107 177 SparseDict=1;
ivan@161 178 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem, SparseDict);
ivan@107 179
idamnjanovic@6 180 % Initialising solver structure
idamnjanovic@6 181 % Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@6 182 % reconstructed and time) to zero values
idamnjanovic@6 183
idamnjanovic@6 184 SMALL.solver(2)=SMALL_init_solver;
idamnjanovic@6 185
idamnjanovic@6 186 % Defining the parameters needed for image denoising
idamnjanovic@6 187
idamnjanovic@6 188 SMALL.solver(2).toolbox='ompsbox';
ivan@107 189 SMALL.solver(2).name='omps2';
ivan@107 190 SMALL.solver(2).param=struct(...
ivan@107 191 'epsilon',Edata,...
ivan@107 192 'maxatoms', maxatoms);
idamnjanovic@6 193
ivan@107 194 % Denoising the image - find the sparse solution in the learned
ivan@107 195 % dictionary for all patches in the image and the end it uses
ivan@107 196 % reconstruction function to reconstruct the patches and put them into a
ivan@107 197 % denoised image
idamnjanovic@6 198
ivan@107 199 SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
idamnjanovic@6 200
ivan@107 201 %%
ivan@107 202 % Use SPAMS Online Dictionary Learning Algorithm
ivan@107 203 % to Learn overcomplete dictionary (Julien Mairal 2009)
ivan@107 204 % (If you have not installed SPAMS please comment the following two cells)
ivan@107 205
ivan@107 206 % Initialising Dictionary structure
ivan@107 207 % Setting Dictionary structure fields (toolbox, name, param, D and time)
ivan@107 208 % to zero values
ivan@107 209
ivan@107 210 SMALL.DL(3)=SMALL_init_DL();
ivan@107 211
ivan@107 212 % Defining fields needed for dictionary learning
ivan@107 213
ivan@107 214 SMALL.DL(3).toolbox = 'SPAMS';
ivan@107 215 SMALL.DL(3).name = 'mexTrainDL';
ivan@107 216
ivan@107 217 % Type 'help mexTrainDL in MATLAB prompt for explanation of parameters.
ivan@107 218
ivan@107 219 SMALL.DL(3).param=struct(...
ivan@107 220 'D', SMALL.Problem.initdict,...
ivan@107 221 'K', SMALL.Problem.p,...
ivan@107 222 'lambda', 2,...
ivan@107 223 'iter', 200,...
ivan@107 224 'mode', 3, ...
ivan@107 225 'modeD', 0);
ivan@107 226
ivan@107 227 % Learn the dictionary
ivan@107 228
ivan@107 229 SMALL.DL(3) = SMALL_learn(SMALL.Problem, SMALL.DL(3));
ivan@107 230
ivan@107 231 % Set SMALL.Problem.A dictionary
ivan@107 232 % (backward compatiblity with SPARCO: solver structure communicate
ivan@107 233 % only with Problem structure, ie no direct communication between DL and
ivan@107 234 % solver structures)
ivan@107 235
ivan@107 236 SMALL.Problem.A = SMALL.DL(3).D;
ivan@107 237
ivan@107 238 % Setting up reconstruction function
ivan@107 239
ivan@161 240 SMALL.Problem.reconstruct = @(x) ImageDenoise_reconstruct(x, SMALL.Problem);
ivan@107 241
ivan@107 242 % Initialising solver structure
ivan@107 243 % Setting solver structure fields (toolbox, name, param, solution,
ivan@107 244 % reconstructed and time) to zero values
ivan@107 245
ivan@107 246 SMALL.solver(3)=SMALL_init_solver;
ivan@107 247
ivan@107 248 % Defining the parameters needed for image denoising
ivan@107 249
ivan@107 250 SMALL.solver(3).toolbox='ompbox';
ivan@107 251 SMALL.solver(3).name='omp2';
ivan@107 252 SMALL.solver(3).param=struct(...
ivan@107 253 'epsilon',Edata,...
ivan@107 254 'maxatoms', maxatoms);
ivan@107 255
ivan@107 256 % Denoising the image - find the sparse solution in the learned
ivan@107 257 % dictionary for all patches in the image and the end it uses
ivan@107 258 % reconstruction function to reconstruct the patches and put them into a
ivan@107 259 % denoised image
ivan@107 260
ivan@107 261 SMALL.solver(3)=SMALL_solve(SMALL.Problem, SMALL.solver(3));
idamnjanovic@6 262
idamnjanovic@6 263 %%
idamnjanovic@6 264 % Plot results and save midi files
idamnjanovic@6 265
idamnjanovic@6 266 % show results %
idamnjanovic@6 267
idamnjanovic@6 268 SMALL_ImgDeNoiseResult(SMALL);