idamnjanovic@43: %% DICTIONARY LEARNING FOR IMAGE DENOISING
idamnjanovic@43: %   This file contains an example of how SMALLbox can be used to test different
idamnjanovic@43: %   dictionary learning techniques in Image Denoising problem.
idamnjanovic@43: %   It calls generateImageDenoiseProblem that will let you to choose image,
idamnjanovic@43: %   add noise and use noisy image to generate training set for dictionary
idamnjanovic@43: %   learning.
idamnjanovic@43: %   Three dictionary learning techniques were compared:
idamnjanovic@43: %   -   KSVD - M. Elad, R. Rubinstein, and M. Zibulevsky, "Efficient
idamnjanovic@43: %              Implementation of the K-SVD Algorithm using Batch Orthogonal
idamnjanovic@43: %              Matching Pursuit", Technical Report - CS, Technion, April 2008.
idamnjanovic@43: %   -   KSVDS - R. Rubinstein, M. Zibulevsky, and M. Elad, "Learning Sparse
idamnjanovic@43: %               Dictionaries for Sparse Signal Approximation", Technical
idamnjanovic@43: %               Report - CS, Technion, June 2009.
idamnjanovic@43: %   -   SPAMS - J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online
idamnjanovic@43: %               Dictionary Learning for Sparse Coding. International
idamnjanovic@43: %               Conference on Machine Learning,Montreal, Canada, 2009
idamnjanovic@43: %
idamnjanovic@43: %
idamnjanovic@43: % Ivan Damnjanovic 2010
idamnjanovic@43: %%
idamnjanovic@43: 
idamnjanovic@43: clear;
idamnjanovic@43: 
idamnjanovic@43: %   If you want to load the image outside of generateImageDenoiseProblem
idamnjanovic@43: %   function uncomment following lines. This can be useful if you want to
idamnjanovic@43: %   denoise more then one image for example.
idamnjanovic@43: 
idamnjanovic@43: % TMPpath=pwd;
idamnjanovic@43: % FS=filesep;
idamnjanovic@43: % [pathstr1, name, ext, versn] = fileparts(which('SMALLboxSetup.m'));
idamnjanovic@43: % cd([pathstr1,FS,'data',FS,'images']);
idamnjanovic@43: % [filename,pathname] = uigetfile({'*.png;'},'Select a file containin pre-calculated notes');
idamnjanovic@43: % [pathstr, name, ext, versn] = fileparts(filename);
idamnjanovic@43: % test_image = imread(filename);
idamnjanovic@43: % test_image = double(test_image);
idamnjanovic@43: % cd(TMPpath);
idamnjanovic@43: % SMALL.Problem.name=name;
idamnjanovic@43: 
idamnjanovic@43: 
idamnjanovic@43: % Defining Image Denoising Problem as Dictionary Learning
idamnjanovic@43: % Problem. As an input we set the number of training patches.
idamnjanovic@43: 
idamnjanovic@43: SMALL.Problem = generateImageDenoiseProblem('', 40000, '','', 20);
idamnjanovic@43: 
idamnjanovic@43: Edata=sqrt(prod(SMALL.Problem.blocksize)) * SMALL.Problem.sigma * SMALL.Problem.gain;
idamnjanovic@43: maxatoms = floor(prod(SMALL.Problem.blocksize)/2);
idamnjanovic@43: %%   Use KSVD Dictionary Learning Algorithm to Learn overcomplete dictionary
idamnjanovic@43: % 
idamnjanovic@43: % %   Initialising Dictionary structure
idamnjanovic@43: % %   Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@43: % %   to zero values
idamnjanovic@43: % 
idamnjanovic@43: % SMALL.DL(1)=SMALL_init_DL();
idamnjanovic@43: % 
idamnjanovic@43: % % Defining the parameters needed for dictionary learning
idamnjanovic@43: % 
idamnjanovic@43: % SMALL.DL(1).toolbox = 'KSVD';
idamnjanovic@43: % SMALL.DL(1).name = 'ksvd';
idamnjanovic@43: % 
idamnjanovic@43: % %   Defining the parameters for KSVD
idamnjanovic@43: % %   In this example we are learning 256 atoms in 20 iterations, so that
idamnjanovic@43: % %   every patch in the training set can be represented with target error in
idamnjanovic@43: % %   L2-norm (EData)
idamnjanovic@43: % %   Type help ksvd in MATLAB prompt for more options.
idamnjanovic@43: % 
idamnjanovic@43: % 
idamnjanovic@43: % SMALL.DL(1).param=struct(...
idamnjanovic@43: %     'Edata', Edata,...
idamnjanovic@43: %     'initdict', SMALL.Problem.initdict,...
idamnjanovic@43: %     'dictsize', SMALL.Problem.p,...
idamnjanovic@43: %     'iternum', 20,...
idamnjanovic@43: %     'memusage', 'high');
idamnjanovic@43: % 
idamnjanovic@43: % %   Learn the dictionary
idamnjanovic@43: % 
idamnjanovic@43: % SMALL.DL(1) = SMALL_learn(SMALL.Problem, SMALL.DL(1));
idamnjanovic@43: %%   Initialising Dictionary structure
idamnjanovic@43: %   Setting Dictionary structure fields (toolbox, name, param, D and time)
idamnjanovic@43: %   to zero values 
idamnjanovic@43: % 
idamnjanovic@43: 
idamnjanovic@43: SMALL.DL(1)=SMALL_init_DL();
idamnjanovic@43: %   Take initial dictonary (overcomplete DCT) to be a final dictionary for
idamnjanovic@43: %   reconstruction
idamnjanovic@43: 
idamnjanovic@43: SMALL.DL(1).D=SMALL.Problem.initdict;
idamnjanovic@43: %%
idamnjanovic@43: 
idamnjanovic@43: %   Set SMALL.Problem.A dictionary
idamnjanovic@43: %   (backward compatiblity with SPARCO: solver structure communicate
idamnjanovic@43: %   only with Problem structure, ie no direct communication between DL and
idamnjanovic@43: %   solver structures)
idamnjanovic@43: SMALL.Problem.A = SMALL.DL(1).D;
idamnjanovic@43: 
idamnjanovic@43: SparseDict=0;
idamnjanovic@43: SMALL.Problem.reconstruct = @(x) ImgDenoise_reconstruct(x, SMALL.Problem, SparseDict);
idamnjanovic@43: 
idamnjanovic@43: %%
idamnjanovic@43: %   Initialising solver structure
idamnjanovic@43: %   Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@43: %   reconstructed and time) to zero values
idamnjanovic@43: 
idamnjanovic@43: 
idamnjanovic@43: SMALL.solver(1)=SMALL_init_solver;
idamnjanovic@43: 
idamnjanovic@43: % Defining the parameters needed for image denoising
idamnjanovic@43: 
idamnjanovic@43: SMALL.solver(1).toolbox='ompbox';
idamnjanovic@43: SMALL.solver(1).name='omp2';
idamnjanovic@43: SMALL.solver(1).param=struct(...
idamnjanovic@43:     'epsilon',Edata,...
idamnjanovic@43:     'maxatoms', maxatoms); 
idamnjanovic@43: 
idamnjanovic@43: %   Denoising the image - SMALL_denoise function is similar to SMALL_solve,
idamnjanovic@43: %   but backward compatible with KSVD definition of denoising
idamnjanovic@43: %   Pay attention that since implicit base dictionary is used, denoising
idamnjanovic@43: %   can be much faster then using explicit dictionary in KSVD example.
idamnjanovic@43: 
idamnjanovic@43: SMALL.solver(1)=SMALL_solve(SMALL.Problem, SMALL.solver(1));
idamnjanovic@43: 
idamnjanovic@43: %%
idamnjanovic@43: %   Initialising solver structure
idamnjanovic@43: %   Setting solver structure fields (toolbox, name, param, solution,
idamnjanovic@43: %   reconstructed and time) to zero values
idamnjanovic@43: lam=2*SMALL.Problem.sigma;%*sqrt(2*log2(size(SMALL.Problem.A,1)))
idamnjanovic@43: for i=1:11
idamnjanovic@43:     lambda(i)=lam+5-(i-1);
idamnjanovic@43: SMALL.DL(2)=SMALL_init_DL();
idamnjanovic@43: i
idamnjanovic@43: %SMALL.Problem.A = SMALL.Problem.initdict;
idamnjanovic@43: SMALL.DL(2).D=SMALL.Problem.initdict;
idamnjanovic@43: SMALL.solver(2)=SMALL_init_solver;
idamnjanovic@43: 
idamnjanovic@43: % Defining the parameters needed for image denoising
idamnjanovic@43: 
idamnjanovic@43: SMALL.solver(2).toolbox='SPAMS';
idamnjanovic@43: SMALL.solver(2).name='mexLasso';
idamnjanovic@43: SMALL.solver(2).param=struct(...
idamnjanovic@43:     'mode', 2, ...
idamnjanovic@43:      'lambda',lambda(i),...
idamnjanovic@43:     'L', maxatoms); 
idamnjanovic@43: 
idamnjanovic@43: %   Denoising the image - SMALL_denoise function is similar to SMALL_solve,
idamnjanovic@43: %   but backward compatible with KSVD definition of denoising
idamnjanovic@43: %   Pay attention that since implicit base dictionary is used, denoising
idamnjanovic@43: %   can be much faster then using explicit dictionary in KSVD example.
idamnjanovic@43: 
idamnjanovic@43: SMALL.solver(2)=SMALL_solve(SMALL.Problem, SMALL.solver(2));
idamnjanovic@43: 
idamnjanovic@43: 
idamnjanovic@43: % show results %
idamnjanovic@43: 
idamnjanovic@43: %SMALL_ImgDeNoiseResult(SMALL);
idamnjanovic@43: 
idamnjanovic@43:     time(1,i) = SMALL.solver(2).time;
idamnjanovic@43:     psnr(1,i) = SMALL.solver(2).reconstructed.psnr;
idamnjanovic@43: end%% show time and psnr %%
idamnjanovic@43: figure('Name', 'SPAMS LAMBDA TEST');
idamnjanovic@43: 
idamnjanovic@43: subplot(1,2,1); plot(lambda, time(1,:), 'ro-');
idamnjanovic@43: title('time vs lambda');
idamnjanovic@43: subplot(1,2,2); plot(lambda, psnr(1,:), 'b*-');
idamnjanovic@43: title('PSNR vs lambda');