annotate examples/Image Denoising/SMALL_ImgDenoise_DL_test_KSVDvsRLSDLAvsTwoStepMOD.m @ 160:e3035d45d014 danieleb

Added support classes
author Daniele Barchiesi <daniele.barchiesi@eecs.qmul.ac.uk>
date Wed, 31 Aug 2011 10:53:10 +0100
parents af307f247ac7
children f42aa8bcb82f
rev   line source
ivan@152 1 %% Dictionary Learning for Image Denoising - KSVD vs Recursive Least Squares
ivan@152 2 %
ivan@152 3 % This file contains an example of how SMALLbox can be used to test different
ivan@152 4 % dictionary learning techniques in Image Denoising problem.
ivan@152 5 % It calls generateImageDenoiseProblem that will let you to choose image,
ivan@152 6 % add noise and use noisy image to generate training set for dictionary
ivan@152 7 % learning.
ivan@152 8 % Two dictionary learning techniques were compared:
ivan@152 9 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
ivan@152 10 % Implementation of the K-SVD Algorithm using Batch Orthogonal
ivan@152 11 % Matching Pursuit", Technical Report - CS, Technion, April 2008.
ivan@152 12 % - RLS-DLA - Skretting, K.; Engan, K.; , "Recursive Least Squares
ivan@152 13 % Dictionary Learning Algorithm," Signal Processing, IEEE Transactions on,
ivan@152 14 % vol.58, no.4, pp.2121-2130, April 2010
ivan@152 15 %
ivan@152 16
ivan@152 17
ivan@152 18 % Centre for Digital Music, Queen Mary, University of London.
ivan@152 19 % This file copyright 2011 Ivan Damnjanovic.
ivan@152 20 %
ivan@152 21 % This program is free software; you can redistribute it and/or
ivan@152 22 % modify it under the terms of the GNU General Public License as
ivan@152 23 % published by the Free Software Foundation; either version 2 of the
ivan@152 24 % License, or (at your option) any later version. See the file
ivan@152 25 % COPYING included with this distribution for more information.
ivan@152 26 %
ivan@152 27 %%
ivan@152 28
ivan@152 29
ivan@152 30
ivan@152 31 % If you want to load the image outside of generateImageDenoiseProblem
ivan@152 32 % function uncomment following lines. This can be useful if you want to
ivan@152 33 % denoise more then one image for example.
ivan@152 34 % Here we are loading test_image.mat that contains structure with 5 images : lena,
ivan@152 35 % barbara,boat, house and peppers.
ivan@152 36 clear;
ivan@152 37 TMPpath=pwd;
ivan@152 38 FS=filesep;
ivan@152 39 [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
ivan@152 40 cd([pathstr1,FS,'data',FS,'images']);
ivan@152 41 load('test_image.mat');
ivan@152 42 cd(TMPpath);
ivan@152 43
ivan@152 44 % Deffining the noise levels that we want to test
ivan@152 45
ivan@152 46 noise_level=[10 20 25 50 100];
ivan@152 47
ivan@152 48 % Here we loop through different noise levels and images
ivan@152 49
ivan@153 50 for noise_ind=4:4
ivan@152 51 for im_num=1:1
ivan@152 52
ivan@152 53 % Defining Image Denoising Problem as Dictionary Learning
ivan@152 54 % Problem. As an input we set the number of training patches.
ivan@152 55
ivan@152 56 SMALL.Problem = generateImageDenoiseProblem(test_image(im_num).i, 40000, '',256, noise_level(noise_ind));
ivan@152 57 SMALL.Problem.name=int2str(im_num);
ivan@152 58
ivan@152 59 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
ivan@152 60 maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
ivan@152 61
ivan@152 62 % results structure is to store all results
ivan@152 63
ivan@152 64 results(noise_ind,im_num).noisy_psnr=SMALL.Problem.noisy_psnr;
ivan@152 65
ivan@152 66 %%
ivan@152 67 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
ivan@152 68
ivan@152 69 % Initialising Dictionary structure
ivan@152 70 % Setting Dictionary structure fields (toolbox, name, param, D and time)
ivan@152 71 % to zero values
ivan@152 72
ivan@152 73 SMALL.DL(1)=SMALL_init_DL();
ivan@152 74
ivan@152 75 % Defining the parameters needed for dictionary learning
ivan@152 76
ivan@152 77 SMALL.DL(1).toolbox = 'KSVD';
ivan@152 78 SMALL.DL(1).name = 'ksvd';
ivan@152 79
ivan@152 80 % Defining the parameters for KSVD
ivan@152 81 % In this example we are learning 256 atoms in 20 iterations, so that
ivan@152 82 % every patch in the training set can be represented with target error in
ivan@152 83 % L2-norm (Edata)
ivan@152 84 % Type help ksvd in MATLAB prompt for more options.
ivan@152 85
ivan@152 86
ivan@152 87 SMALL.DL(1).param=struct(...
ivan@152 88 'Edata', Edata,...
ivan@152 89 'initdict', SMALL.Problem.initdict,...
ivan@152 90 'dictsize', SMALL.Problem.p,...
ivan@152 91 'exact', 1, ...
ivan@152 92 'iternum', 20,...
ivan@152 93 'memusage', 'high');
ivan@152 94
ivan@152 95 % Learn the dictionary
ivan@152 96
ivan@152 97 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
ivan@152 98
ivan@152 99 % Set SMALL.Problem.A dictionary
ivan@152 100 % (backward compatiblity with SPARCO: solver structure communicate
ivan@152 101 % only with Problem structure, ie no direct communication between DL and
ivan@152 102 % solver structures)
ivan@152 103
ivan@152 104 SMALL.Problem.A = SMALL.DL(1).D;
ivan@152 105 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
ivan@152 106
ivan@152 107 %%
ivan@152 108 % Initialising solver structure
ivan@152 109 % Setting solver structure fields (toolbox, name, param, solution,
ivan@152 110 % reconstructed and time) to zero values
ivan@152 111
ivan@152 112 SMALL.solver(1)=SMALL_init_solver;
ivan@152 113
ivan@152 114 % Defining the parameters needed for image denoising
ivan@152 115
ivan@152 116 SMALL.solver(1).toolbox='ompbox';
ivan@152 117 SMALL.solver(1).name='omp2';
ivan@152 118 SMALL.solver(1).param=struct(...
ivan@152 119 'epsilon',Edata,...
ivan@152 120 'maxatoms', maxatoms);
ivan@152 121
ivan@152 122 % Denoising the image - find the sparse solution in the learned
ivan@152 123 % dictionary for all patches in the image and the end it uses
ivan@152 124 % reconstruction function to reconstruct the patches and put them into a
ivan@152 125 % denoised image
ivan@152 126
ivan@152 127 SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
ivan@152 128
ivan@152 129 % Show PSNR after reconstruction
ivan@152 130
ivan@152 131 SMALL.solver(1).reconstructed.psnr
ivan@152 132
ivan@152 133 %%
ivan@152 134 % For comparison purposes we will denoise image with overcomplete DCT
ivan@152 135 % here
ivan@152 136 % Set SMALL.Problem.A dictionary to be oDCT (i.e. Problem.initdict -
ivan@152 137 % since initial dictionaruy is already set to be oDCT when generating the
ivan@152 138 % denoising problem
ivan@152 139
ivan@152 140
ivan@152 141 % Initialising solver structure
ivan@152 142 % Setting solver structure fields (toolbox, name, param, solution,
ivan@152 143 % reconstructed and time) to zero values
ivan@152 144
ivan@152 145 SMALL.solver(2)=SMALL_init_solver;
ivan@152 146
ivan@152 147 % Defining the parameters needed for image denoising
ivan@152 148
ivan@152 149 SMALL.solver(2).toolbox='ompbox';
ivan@152 150 SMALL.solver(2).name='omp2';
ivan@152 151 SMALL.solver(2).param=struct(...
ivan@152 152 'epsilon',Edata,...
ivan@152 153 'maxatoms', maxatoms);
ivan@152 154
ivan@152 155 % Initialising Dictionary structure
ivan@152 156 % Setting Dictionary structure fields (toolbox, name, param, D and time)
ivan@152 157 % to zero values
ivan@152 158
ivan@152 159 SMALL.DL(2)=SMALL_init_DL('TwoStepDL', 'MOD', '', 1);
ivan@152 160
ivan@152 161
ivan@152 162 % Defining the parameters for MOD
ivan@152 163 % In this example we are learning 256 atoms in 20 iterations, so that
ivan@152 164 % every patch in the training set can be represented with target error in
ivan@152 165 % L2-norm (EData)
ivan@152 166 % Type help ksvd in MATLAB prompt for more options.
ivan@152 167
ivan@152 168
ivan@152 169 SMALL.DL(2).param=struct(...
ivan@152 170 'solver', SMALL.solver(2),...
ivan@152 171 'initdict', SMALL.Problem.initdict,...
ivan@152 172 'dictsize', SMALL.Problem.p,...
ivan@152 173 'iternum', 40,...
ivan@152 174 'show_dict', 1);
ivan@152 175
ivan@152 176 % Learn the dictionary
ivan@152 177
ivan@152 178 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2));
ivan@152 179
ivan@152 180 % Set SMALL.Problem.A dictionary
ivan@152 181 % (backward compatiblity with SPARCO: solver structure communicate
ivan@152 182 % only with Problem structure, ie no direct communication between DL and
ivan@152 183 % solver structures)
ivan@152 184
ivan@152 185 SMALL.Problem.A = SMALL.DL(2).D;
ivan@152 186 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
ivan@152 187
ivan@152 188 % Denoising the image - find the sparse solution in the learned
ivan@152 189 % dictionary for all patches in the image and the end it uses
ivan@152 190 % reconstruction function to reconstruct the patches and put them into a
ivan@152 191 % denoised image
ivan@152 192
ivan@152 193 SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
ivan@152 194
ivan@152 195 %%
ivan@152 196 % In the b1 field all patches from the image are stored. For RLS-DLA we
ivan@152 197 % will first exclude all the patches that have l2 norm smaller then
ivan@152 198 % threshold and then take min(40000, number_of_remaining_patches) in
ivan@152 199 % ascending order as our training set (SMALL.Problem.b)
ivan@152 200
ivan@152 201 X=SMALL.Problem.b1;
ivan@152 202 X_norm=sqrt(sum(X.^2, 1));
ivan@152 203 [X_norm_sort, p]=sort(X_norm);
ivan@152 204 p1=p(X_norm_sort>Edata);
ivan@152 205 if size(p1,2)>40000
ivan@152 206 p2 = randperm(size(p1,2));
ivan@152 207 p2=sort(p2(1:40000));
ivan@152 208 size(p2,2)
ivan@152 209 SMALL.Problem.b=X(:,p1(p2));
ivan@152 210 else
ivan@152 211 size(p1,2)
ivan@152 212 SMALL.Problem.b=X(:,p1);
ivan@152 213
ivan@152 214 end
ivan@152 215
ivan@152 216 % Forgetting factor for RLS-DLA algorithm, in this case we are using
ivan@152 217 % fixed value
ivan@152 218
ivan@152 219 lambda=0.9998
ivan@152 220
ivan@152 221 % Use Recursive Least Squares
ivan@152 222 % to Learn overcomplete dictionary
ivan@152 223
ivan@152 224 % Initialising Dictionary structure
ivan@152 225 % Setting Dictionary structure fields (toolbox, name, param, D and time)
ivan@152 226 % to zero values
ivan@152 227
ivan@152 228 SMALL.DL(3)=SMALL_init_DL();
ivan@152 229
ivan@152 230 % Defining fields needed for dictionary learning
ivan@152 231
ivan@152 232 SMALL.DL(3).toolbox = 'SMALL';
ivan@152 233 SMALL.DL(3).name = 'SMALL_rlsdla';
ivan@152 234 SMALL.DL(3).param=struct(...
ivan@152 235 'Edata', Edata,...
ivan@152 236 'initdict', SMALL.Problem.initdict,...
ivan@152 237 'dictsize', SMALL.Problem.p,...
ivan@152 238 'forgettingMode', 'FIX',...
ivan@152 239 'forgettingFactor', lambda,...
ivan@152 240 'show_dict', 1000);
ivan@152 241
ivan@152 242
ivan@152 243 SMALL.DL(3) = SMALL_learn(SMALL.Problem, SMALL.DL(3));
ivan@152 244
ivan@152 245 % Initialising solver structure
ivan@152 246 % Setting solver structure fields (toolbox, name, param, solution,
ivan@152 247 % reconstructed and time) to zero values
ivan@152 248
ivan@152 249 SMALL.Problem.A = SMALL.DL(3).D;
ivan@152 250 SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
ivan@152 251
ivan@152 252 SMALL.solver(3)=SMALL_init_solver;
ivan@152 253
ivan@152 254 % Defining the parameters needed for image denoising
ivan@152 255
ivan@152 256 SMALL.solver(3).toolbox='ompbox';
ivan@152 257 SMALL.solver(3).name='omp2';
ivan@152 258 SMALL.solver(3).param=struct(...
ivan@152 259 'epsilon',Edata,...
ivan@152 260 'maxatoms', maxatoms);
ivan@152 261
ivan@152 262
ivan@152 263 SMALL.solver(3)=SMALL_solve(SMALL.Problem, SMALL.solver(3));
ivan@152 264
ivan@152 265 SMALL.solver(3).reconstructed.psnr
ivan@152 266
ivan@152 267
ivan@152 268 % show results %
ivan@152 269
ivan@152 270 SMALL_ImgDeNoiseResult(SMALL);
ivan@152 271
ivan@152 272 results(noise_ind,im_num).psnr.ksvd=SMALL.solver(1).reconstructed.psnr;
ivan@152 273 results(noise_ind,im_num).psnr.odct=SMALL.solver(2).reconstructed.psnr;
ivan@152 274 results(noise_ind,im_num).psnr.rlsdla=SMALL.solver(3).reconstructed.psnr;
ivan@152 275 results(noise_ind,im_num).vmrse.ksvd=SMALL.solver(1).reconstructed.vmrse;
ivan@152 276 results(noise_ind,im_num).vmrse.odct=SMALL.solver(2).reconstructed.vmrse;
ivan@152 277 results(noise_ind,im_num).vmrse.rlsdla=SMALL.solver(3).reconstructed.vmrse;
ivan@152 278 results(noise_ind,im_num).ssim.ksvd=SMALL.solver(1).reconstructed.ssim;
ivan@152 279 results(noise_ind,im_num).ssim.odct=SMALL.solver(2).reconstructed.ssim;
ivan@152 280 results(noise_ind,im_num).ssim.rlsdla=SMALL.solver(3).reconstructed.ssim;
ivan@152 281
ivan@152 282 results(noise_ind,im_num).time.ksvd=SMALL.solver(1).time+SMALL.DL(1).time;
ivan@152 283 results(noise_ind,im_num).time.rlsdla.time=SMALL.solver(3).time+SMALL.DL(3).time;
ivan@152 284 clear SMALL;
ivan@152 285 end
ivan@152 286 end
ivan@152 287 % save results.mat results