annotate examples/Image Denoising/SMALL_ImgDenoise_DL_test_Training_size.m @ 6:f72603404233

(none)
author idamnjanovic
date Mon, 22 Mar 2010 10:45:01 +0000
parents
children 79e1d62f0115
rev   line source
idamnjanovic@6 1 %% DICTIONARY LEARNING FOR IMAGE DENOISING
idamnjanovic@6 2 % This file contains an example of how SMALLbox can be used to test different
idamnjanovic@6 3 % dictionary learning techniques in Image Denoising problem.
idamnjanovic@6 4 % It calls generateImageDenoiseProblem that will let you to choose image,
idamnjanovic@6 5 % add noise and use noisy image to generate training set for dictionary
idamnjanovic@6 6 % learning.
idamnjanovic@6 7 % We tested time and psnr for two dictionary learning techniques. This
idamnjanovic@6 8 % example does not represnt any extensive testing. The aim of this
idamnjanovic@6 9 % example is just to show how SMALL structure can be used for testing.
idamnjanovic@6 10 %
idamnjanovic@6 11 % Two dictionary learning techniques were compared:
idamnjanovic@6 12 % - KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
idamnjanovic@6 13 % Implementation of the K-SVD Algorithm using Batch Orthogonal
idamnjanovic@6 14 % Matching Pursuit", Technical Report - CS, Technion, April 2008.
idamnjanovic@6 15 % - SPAMS - J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online
idamnjanovic@6 16 % Dictionary Learning for Sparse Coding. International
idamnjanovic@6 17 % Conference on Machine Learning,Montreal, Canada, 2009
idamnjanovic@6 18 %
idamnjanovic@6 19 %
idamnjanovic@6 20 % Ivan Damnjanovic 2010
idamnjanovic@6 21 %%
idamnjanovic@6 22
idamnjanovic@6 23 clear all;
idamnjanovic@6 24
idamnjanovic@6 25 %% Load an image
idamnjanovic@6 26 TMPpath=pwd;
idamnjanovic@6 27 FS=filesep;
idamnjanovic@6 28 [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
idamnjanovic@6 29 cd([pathstr1,FS,'data',FS,'images']);
idamnjanovic@6 30 [filename,pathname] = uigetfile({'*.png;'},'Select a file containin pre-calculated notes');
idamnjanovic@6 31 [pathstr, name, ext, versn] = fileparts(filename);
idamnjanovic@6 32 test_image = imread(filename);
idamnjanovic@6 33 test_image = double(test_image);
idamnjanovic@6 34 cd(TMPpath);
idamnjanovic@6 35
idamnjanovic@6 36 % number of different values we want to test
idamnjanovic@6 37 n =5;
idamnjanovic@6 38 step = floor((size(test_image,1)-8+1)*(size(test_image,2)-8+1)/n);
idamnjanovic@6 39 Training_size=zeros(1,n);
idamnjanovic@6 40 time = zeros(2,n);
idamnjanovic@6 41 psnr = zeros(2,n);
idamnjanovic@6 42 for i=1:n
idamnjanovic@6 43
idamnjanovic@6 44 % Here we want to test time spent and quality of denoising for
idamnjanovic@6 45 % different sizes of training sample.
idamnjanovic@6 46 Training_size(i)=i*step;
idamnjanovic@6 47
idamnjanovic@6 48 SMALL.Problem = generateImageDenoiseProblem(test_image,Training_size(i));
idamnjanovic@6 49 SMALL.Problem.name=name;
idamnjanovic@6 50 %%
idamnjanovic@6 51 % Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
idamnjanovic@6 52
idamnjanovic@6 53 % Initialising Dictionary structure
idamnjanovic@6 54 % Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@6 55 % to zero values
idamnjanovic@6 56
idamnjanovic@6 57 SMALL.DL(1)=SMALL_init_DL();
idamnjanovic@6 58
idamnjanovic@6 59 % Defining the parameters needed for dictionary learning
idamnjanovic@6 60
idamnjanovic@6 61 SMALL.DL(1).toolbox = 'KSVD';
idamnjanovic@6 62 SMALL.DL(1).name = 'ksvd';
idamnjanovic@6 63
idamnjanovic@6 64 % Defining the parameters for KSVD
idamnjanovic@6 65 % In this example we are learning 256 atoms in 20 iterations, so that
idamnjanovic@6 66 % every patch in the training set can be represented with target error in
idamnjanovic@6 67 % L2-norm (EData)
idamnjanovic@6 68 % Type help ksvd in MATLAB prompt for more options.
idamnjanovic@6 69
idamnjanovic@6 70 Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
idamnjanovic@6 71 SMALL.DL(1).param=struct(...
idamnjanovic@6 72 'Edata', Edata,...
idamnjanovic@6 73 'initdict', SMALL.Problem.initdict,...
idamnjanovic@6 74 'dictsize', SMALL.Problem.p,...
idamnjanovic@6 75 'iternum', 20,...
idamnjanovic@6 76 'memusage', 'high');
idamnjanovic@6 77
idamnjanovic@6 78 % Learn the dictionary
idamnjanovic@6 79
idamnjanovic@6 80 SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
idamnjanovic@6 81
idamnjanovic@6 82 % Set SMALL.Problem.A dictionary
idamnjanovic@6 83 % (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@6 84 % only with Problem structure, ie no direct communication between DL and
idamnjanovic@6 85 % solver structures)
idamnjanovic@6 86
idamnjanovic@6 87 SMALL.Problem.A = SMALL.DL(1).D;
idamnjanovic@6 88
idamnjanovic@6 89
idamnjanovic@6 90 %%
idamnjanovic@6 91 % Initialising solver structure
idamnjanovic@6 92 % Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@6 93 % reconstructed and time) to zero values
idamnjanovic@6 94
idamnjanovic@6 95
idamnjanovic@6 96 SMALL.solver(1)=SMALL_init_solver;
idamnjanovic@6 97
idamnjanovic@6 98 % Defining the parameters needed for denoising
idamnjanovic@6 99
idamnjanovic@6 100 SMALL.solver(1).toolbox='ompbox';
idamnjanovic@6 101 SMALL.solver(1).name='ompdenoise';
idamnjanovic@6 102
idamnjanovic@6 103 % Denoising the image - SMALL_denoise function is similar to SMALL_solve,
idamnjanovic@6 104 % but backward compatible with KSVD definition of denoising
idamnjanovic@6 105
idamnjanovic@6 106 SMALL.solver(1)=SMALL_denoise(SMALL.Problem, SMALL.solver(1));
idamnjanovic@6 107
idamnjanovic@6 108 %%
idamnjanovic@6 109 % Use SPAMS Online Dictionary Learning Algorithm
idamnjanovic@6 110 % to Learn overcomplete dictionary (Julien Mairal 2009)
idamnjanovic@6 111 % (If you have not installed SPAMS please comment the following two cells)
idamnjanovic@6 112
idamnjanovic@6 113 % Initialising Dictionary structure
idamnjanovic@6 114 % Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@6 115 % to zero values
idamnjanovic@6 116
idamnjanovic@6 117 SMALL.DL(2)=SMALL_init_DL();
idamnjanovic@6 118
idamnjanovic@6 119 % Defining fields needed for dictionary learning
idamnjanovic@6 120
idamnjanovic@6 121 SMALL.DL(2).toolbox = 'SPAMS';
idamnjanovic@6 122 SMALL.DL(2).name = 'mexTrainDL';
idamnjanovic@6 123
idamnjanovic@6 124 % Type 'help mexTrainDL in MATLAB prompt for explanation of parameters.
idamnjanovic@6 125
idamnjanovic@6 126 SMALL.DL(2).param=struct(...
idamnjanovic@6 127 'D', SMALL.Problem.initdict,...
idamnjanovic@6 128 'K', SMALL.Problem.p,...
idamnjanovic@6 129 'lambda', 2,...
idamnjanovic@6 130 'iter', 300,...
idamnjanovic@6 131 'mode', 3,...
idamnjanovic@6 132 'modeD', 0 );
idamnjanovic@6 133
idamnjanovic@6 134 % Learn the dictionary
idamnjanovic@6 135
idamnjanovic@6 136 SMALL.DL(2) = SMALL_learn(SMALL.Problem, SMALL.DL(2));
idamnjanovic@6 137
idamnjanovic@6 138 % Set SMALL.Problem.A dictionary
idamnjanovic@6 139 % (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@6 140 % only with Problem structure, ie no direct communication between DL and
idamnjanovic@6 141 % solver structures)
idamnjanovic@6 142
idamnjanovic@6 143 SMALL.Problem.A = SMALL.DL(2).D;
idamnjanovic@6 144
idamnjanovic@6 145
idamnjanovic@6 146 %%
idamnjanovic@6 147 % Initialising solver structure
idamnjanovic@6 148 % Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@6 149 % reconstructed and time) to zero values
idamnjanovic@6 150
idamnjanovic@6 151 SMALL.solver(2)=SMALL_init_solver;
idamnjanovic@6 152
idamnjanovic@6 153 % Defining the parameters needed for denoising
idamnjanovic@6 154
idamnjanovic@6 155 SMALL.solver(2).toolbox='ompbox';
idamnjanovic@6 156 SMALL.solver(2).name='ompdenoise';
idamnjanovic@6 157
idamnjanovic@6 158 % Denoising the image - SMALL_denoise function is similar to SMALL_solve,
idamnjanovic@6 159 % but backward compatible with KSVD definition of denoising
idamnjanovic@6 160
idamnjanovic@6 161 SMALL.solver(2)=SMALL_denoise(SMALL.Problem, SMALL.solver(2));
idamnjanovic@6 162
idamnjanovic@6 163
idamnjanovic@6 164
idamnjanovic@6 165 %% show results %%
idamnjanovic@6 166 % This will show denoised images and dictionaries for all training sets.
idamnjanovic@6 167 % If you are not interested to see them and do not want clutter your
idamnjanovic@6 168 % screen comment following line
idamnjanovic@6 169
idamnjanovic@6 170 SMALL_ImgDeNoiseResult(SMALL);
idamnjanovic@6 171
idamnjanovic@6 172 time(1,i) = SMALL.DL(1).time;
idamnjanovic@6 173 psnr(1,i) = SMALL.solver(1).reconstructed.psnr;
idamnjanovic@6 174
idamnjanovic@6 175 time(2,i) = SMALL.DL(2).time;
idamnjanovic@6 176 psnr(2,i) = SMALL.solver(2).reconstructed.psnr;
idamnjanovic@6 177
idamnjanovic@6 178 clear SMALL
idamnjanovic@6 179 end
idamnjanovic@6 180
idamnjanovic@6 181 %% show time and psnr %%
idamnjanovic@6 182 figure('Name', 'KSVD vs SPAMS');
idamnjanovic@6 183
idamnjanovic@6 184 subplot(1,2,1); plot(Training_size, time(1,:), 'ro-', Training_size, time(2,:), 'b*-');
idamnjanovic@6 185 legend('KSVD','SPAMS',0);
idamnjanovic@6 186 title('Time vs Training size');
idamnjanovic@6 187 subplot(1,2,2); plot(Training_size, psnr(1,:), 'ro-', Training_size, psnr(2,:), 'b*-');
idamnjanovic@6 188 legend('KSVD','SPAMS',0);
idamnjanovic@6 189 title('PSNR vs Training size');