idamnjanovic@42: %% DICTIONARY LEARNING FOR IMAGE DENOISING
idamnjanovic@42: %   This file contains an example of how SMALLbox can be used to test different
idamnjanovic@42: %   dictionary learning techniques in Image Denoising problem.
idamnjanovic@42: %   It calls generateImageDenoiseProblem that will let you to choose image,
idamnjanovic@42: %   add noise and use noisy image to generate training set for dictionary
idamnjanovic@42: %   learning.
maria@83: %   Two dictionary learning techniques were compared:
idamnjanovic@42: %   -   KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
idamnjanovic@42: %              Implementation of the K-SVD Algorithm using Batch Orthogonal
idamnjanovic@42: %              Matching Pursuit", Technical Report - CS, Technion, April 2008.
maria@83: %   -   RLS-DLA - Skretting, K.; Engan, K.; , "Recursive Least Squares
maria@83: %       Dictionary Learning Algorithm," Signal Processing, IEEE Transactions on,
maria@83: %       vol.58, no.4, pp.2121-2130, April 2010
idamnjanovic@42: %
maria@83: 
maria@83: 
maria@83: %   Centre for Digital Music, Queen Mary, University of London.
maria@83: %   This file copyright 2011 Ivan Damnjanovic.
idamnjanovic@42: %
maria@83: %   This program is free software; you can redistribute it and/or
maria@83: %   modify it under the terms of the GNU General Public License as
maria@83: %   published by the Free Software Foundation; either version 2 of the
maria@83: %   License, or (at your option) any later version.  See the file
maria@83: %   COPYING included with this distribution for more information.
maria@83: %   
idamnjanovic@42: %%
idamnjanovic@42: 
idamnjanovic@42: 
idamnjanovic@42: 
idamnjanovic@42: %   If you want to load the image outside of generateImageDenoiseProblem
idamnjanovic@42: %   function uncomment following lines. This can be useful if you want to
idamnjanovic@42: %   denoise more then one image for example.
maria@83: %   Here we are loading test_image.mat that contains structure with 5 images : lena,
maria@83: %   barbara,boat, house and peppers.
idamnjanovic@42: clear;
idamnjanovic@42: TMPpath=pwd;
idamnjanovic@42: FS=filesep;
idamnjanovic@42: [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
idamnjanovic@42: cd([pathstr1,FS,'data',FS,'images']);
idamnjanovic@42: load('test_image.mat');
ivan@77: cd(TMPpath);
maria@83: 
maria@83: %   Deffining the noise levels that we want to test
idamnjanovic@42: 
idamnjanovic@42: noise_level=[10 20 25 50 100];
maria@83: 
maria@83: %   Here we loop through different noise levels and images 
maria@83: 
maria@83: for noise_ind=2:2
maria@83: for im_num=1:1
maria@83: 
idamnjanovic@42: % Defining Image Denoising Problem as Dictionary Learning
idamnjanovic@42: % Problem. As an input we set the number of training patches.
maria@83: 
idamnjanovic@65: SMALL.Problem = generateImageDenoiseProblem(test_image(im_num).i, 40000, '',256, noise_level(noise_ind));
ivan@77: SMALL.Problem.name=int2str(im_num);
ivan@84: %   results structure is to store all results
idamnjanovic@42: 
idamnjanovic@42: results(noise_ind,im_num).noisy_psnr=SMALL.Problem.noisy_psnr;
idamnjanovic@42: 
idamnjanovic@42: %%
idamnjanovic@42: %   Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
idamnjanovic@42: 
idamnjanovic@42: %   Initialising Dictionary structure
idamnjanovic@42: %   Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@42: %   to zero values
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(1)=SMALL_init_DL();
idamnjanovic@42: 
idamnjanovic@42: % Defining the parameters needed for dictionary learning
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(1).toolbox = 'KSVD';
idamnjanovic@42: SMALL.DL(1).name = 'ksvd';
idamnjanovic@42: 
idamnjanovic@42: %   Defining the parameters for KSVD
idamnjanovic@42: %   In this example we are learning 256 atoms in 20 iterations, so that
idamnjanovic@42: %   every patch in the training set can be represented with target error in
idamnjanovic@42: %   L2-norm (EData)
idamnjanovic@42: %   Type help ksvd in MATLAB prompt for more options.
idamnjanovic@42: 
idamnjanovic@42: Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
idamnjanovic@42: maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
idamnjanovic@42: SMALL.DL(1).param=struct(...
idamnjanovic@42:     'Edata', Edata,...
idamnjanovic@42:     'initdict', SMALL.Problem.initdict,...
idamnjanovic@42:     'dictsize', SMALL.Problem.p,...
idamnjanovic@42:     'exact', 1, ...
idamnjanovic@42:     'iternum', 20,...
idamnjanovic@42:     'memusage', 'high');
idamnjanovic@42: 
idamnjanovic@42: %   Learn the dictionary
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
idamnjanovic@42: 
idamnjanovic@42: %   Set SMALL.Problem.A dictionary
idamnjanovic@42: %   (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@42: %   only with Problem structure, ie no direct communication between DL and
idamnjanovic@42: %   solver structures)
idamnjanovic@42: 
idamnjanovic@42: SMALL.Problem.A = SMALL.DL(1).D;
idamnjanovic@42: SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
idamnjanovic@42: 
idamnjanovic@42: %%
idamnjanovic@42: %   Initialising solver structure
idamnjanovic@42: %   Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@42: %   reconstructed and time) to zero values
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(1)=SMALL_init_solver;
idamnjanovic@42: 
idamnjanovic@42: % Defining the parameters needed for image denoising
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(1).toolbox='ompbox';
idamnjanovic@42: SMALL.solver(1).name='omp2';
idamnjanovic@42: SMALL.solver(1).param=struct(...
idamnjanovic@42:     'epsilon',Edata,...
idamnjanovic@42:     'maxatoms', maxatoms); 
idamnjanovic@42: 
ivan@84: %   Denoising the image - find the sparse solution in the learned
ivan@84: %   dictionary for all patches in the image and the end it uses
ivan@84: %   reconstruction function to reconstruct the patches and put them into a
ivan@84: %   denoised image
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
ivan@84: 
ivan@84: %   Show PSNR after reconstruction
ivan@84: 
idamnjanovic@42: SMALL.solver(1).reconstructed.psnr
ivan@84: 
idamnjanovic@42: %%
ivan@84: %   For comparison purposes we will denoise image with overcomplete DCT
ivan@84: %   here
ivan@84: %   Set SMALL.Problem.A dictionary to be oDCT (i.e. Problem.initdict -
ivan@84: %   since initial dictionaruy is already set to be oDCT when generating the
ivan@84: %   denoising problem
idamnjanovic@42: 
idamnjanovic@42: SMALL.Problem.A = SMALL.Problem.initdict;
idamnjanovic@42: SMALL.DL(2).D=SMALL.Problem.initdict;
ivan@84: 
ivan@84: %   Setting up reconstruction function
ivan@84: 
idamnjanovic@42: SparseDict=0;
idamnjanovic@42: SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem, SparseDict);
idamnjanovic@42: 
idamnjanovic@42: %   Initialising solver structure
idamnjanovic@42: %   Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@42: %   reconstructed and time) to zero values
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(2)=SMALL_init_solver;
idamnjanovic@42: 
idamnjanovic@42: % Defining the parameters needed for image denoising
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(2).toolbox='ompbox';
idamnjanovic@42: SMALL.solver(2).name='omp2';
idamnjanovic@42: SMALL.solver(2).param=struct(...
idamnjanovic@42:     'epsilon',Edata,...
idamnjanovic@42:     'maxatoms', maxatoms); 
idamnjanovic@42: 
ivan@84: %   Denoising the image - find the sparse solution in the learned
ivan@84: %   dictionary for all patches in the image and the end it uses
ivan@84: %   reconstruction function to reconstruct the patches and put them into a
ivan@84: %   denoised image
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
idamnjanovic@42: %%
ivan@84: % In the b1 field all patches from the image are stored. For RLS-DLA we
ivan@84: % will first exclude all the patches that have l2 norm smaller then
ivan@84: % threshold and then take min(40000, number_of_remaining_patches) in
ivan@84: % ascending order as our training set (SMALL.Problem.b)
idamnjanovic@42: 
idamnjanovic@42: X=SMALL.Problem.b1;
idamnjanovic@42: X_norm=sqrt(sum(X.^2, 1));
idamnjanovic@42: [X_norm_sort, p]=sort(X_norm);
idamnjanovic@42: p1=p(X_norm_sort>Edata);
maria@83: if size(p1,2)>40000
idamnjanovic@42:     p2 = randperm(size(p1,2));
idamnjanovic@42:     p2=sort(p2(1:40000));
idamnjanovic@42:     size(p2,2)
idamnjanovic@42:     SMALL.Problem.b=X(:,p1(p2));
idamnjanovic@42: else 
idamnjanovic@42:     size(p1,2)
idamnjanovic@42:     SMALL.Problem.b=X(:,p1);
idamnjanovic@42: 
idamnjanovic@42: end
idamnjanovic@42: 
ivan@84: %   Forgetting factor for RLS-DLA algorithm, in this case we are using
ivan@84: %   fixed value
ivan@84: 
idamnjanovic@42: lambda=0.9998
idamnjanovic@42: 
idamnjanovic@42: %   Use Recursive Least Squares
idamnjanovic@42: %   to Learn overcomplete dictionary 
idamnjanovic@42: 
idamnjanovic@42: %   Initialising Dictionary structure
idamnjanovic@42: %   Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@42: %   to zero values
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(3)=SMALL_init_DL();
idamnjanovic@42: 
idamnjanovic@42: %   Defining fields needed for dictionary learning
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(3).toolbox = 'SMALL';
idamnjanovic@42: SMALL.DL(3).name = 'SMALL_rlsdla';
idamnjanovic@42: SMALL.DL(3).param=struct(...
idamnjanovic@42:     'Edata', Edata,...
idamnjanovic@42:     'initdict', SMALL.Problem.initdict,...
idamnjanovic@42:     'dictsize', SMALL.Problem.p,...
idamnjanovic@42:     'forgettingMode', 'FIX',...
ivan@85:     'forgettingFactor', lambda,...
ivan@85:     'show_dict', 500);
idamnjanovic@42: 
idamnjanovic@42: 
idamnjanovic@42: SMALL.DL(3) = SMALL_learn(SMALL.Problem, SMALL.DL(3));
idamnjanovic@42: 
idamnjanovic@42: %   Initialising solver structure
idamnjanovic@42: %   Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@42: %   reconstructed and time) to zero values
idamnjanovic@42: 
ivan@84: SMALL.Problem.A = SMALL.DL(3).D;
ivan@84: SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem);
ivan@84: 
idamnjanovic@42: SMALL.solver(3)=SMALL_init_solver;
idamnjanovic@42: 
idamnjanovic@42: % Defining the parameters needed for image denoising
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(3).toolbox='ompbox';
idamnjanovic@42: SMALL.solver(3).name='omp2';
idamnjanovic@42: SMALL.solver(3).param=struct(...
idamnjanovic@42:     'epsilon',Edata,...
idamnjanovic@42:     'maxatoms', maxatoms); 
idamnjanovic@42: 
idamnjanovic@42: 
idamnjanovic@42: SMALL.solver(3)=SMALL_solve(SMALL.Problem, SMALL.solver(3));
ivan@84: 
idamnjanovic@42: SMALL.solver(3).reconstructed.psnr
ivan@84: 
ivan@84: 
idamnjanovic@42: % show results %
idamnjanovic@42: 
idamnjanovic@42: SMALL_ImgDeNoiseResult(SMALL);
ivan@84: 
idamnjanovic@42: results(noise_ind,im_num).psnr.ksvd=SMALL.solver(1).reconstructed.psnr;
idamnjanovic@42: results(noise_ind,im_num).psnr.odct=SMALL.solver(2).reconstructed.psnr;
idamnjanovic@42: results(noise_ind,im_num).psnr.rlsdla=SMALL.solver(3).reconstructed.psnr;
idamnjanovic@42: results(noise_ind,im_num).vmrse.ksvd=SMALL.solver(1).reconstructed.vmrse;
idamnjanovic@42: results(noise_ind,im_num).vmrse.odct=SMALL.solver(2).reconstructed.vmrse;
idamnjanovic@42: results(noise_ind,im_num).vmrse.rlsdla=SMALL.solver(3).reconstructed.vmrse;
idamnjanovic@42: results(noise_ind,im_num).ssim.ksvd=SMALL.solver(1).reconstructed.ssim;
idamnjanovic@42: results(noise_ind,im_num).ssim.odct=SMALL.solver(2).reconstructed.ssim;
idamnjanovic@42: results(noise_ind,im_num).ssim.rlsdla=SMALL.solver(3).reconstructed.ssim;
idamnjanovic@42: 
idamnjanovic@42: results(noise_ind,im_num).time.ksvd=SMALL.solver(1).time+SMALL.DL(1).time;
idamnjanovic@42: results(noise_ind,im_num).time.rlsdla.time=SMALL.solver(3).time+SMALL.DL(3).time;
idamnjanovic@42: %clear SMALL;
idamnjanovic@42: end
idamnjanovic@42: end
idamnjanovic@42: save results.mat results