comparison examples/SMALL_test_coherence2.m @ 169:290cca7d3469 danieleb

Added dictionary decorrelation functions and test script for ICASSP paper.

+clc, clear, close all
+
+%% Parameteres
+% Dictionary learning parameters
+toolbox   = 'TwoStepDL';							%dictionary learning toolbox
+dicUpdate = {'ksvd'};								%dictionary learning updates
+dicDecorr = {'none','mailhe','tropp','barchiesi'};	%dictionary decorrelation methods
+minCoherence = linspace(0.1,1,1);					%coherence levels
+minCoherence = 0.4;
+%dicDecorr = {'barchiesi'};
+
+iterNum   = 20;				%number of iterations
+epsilon   = 1e-6;			%tolerance level
+dictSize  = 512;			%number of atoms in the dictionary
+percActiveAtoms = 5;		%percentage of active atoms
+
+% Test signal parameters
+signal    = audio('music03_16kHz.wav'); %audio signal
+blockSize = 256;						%size of audio frames
+overlap   = 0.5;						%overlap between consecutive frames
+
+
+% Dependent parameters
+nActiveAtoms = fix(blockSize/100*percActiveAtoms); %number of active atoms
+
+% Initial dictionaries
+dctDict = dictionary('dct',blockSize,dictSize);
+dctDict = dctDict.phi;
+gaborParam = struct('N',blockSize,'redundancyFactor',2,'wd',@rectwin);
+gaborDict = Gabor_Dictionary(gaborParam);
+initDicts = {[],dctDict,gaborDict};
+initDicts = {[]};
+
+%% Generate audio approximation problem
+signal			 = buffer(signal,blockSize,blockSize*overlap,@rectwin);
+SMALL.Problem.b  = signal.S;
+SMALL.Problem.b1 = SMALL.Problem.b; % copy signals from training set b to test set b1 (needed for later functions)
+
+% omp2 sparse representation solver
+ompParam = struct('X',SMALL.Problem.b,'epsilon',epsilon,'maxatoms',nActiveAtoms); %parameters
+solver	 = SMALL_init_solver('ompbox','omp2',ompParam,false); %solver structure
+
+
+%% Test
+nDicUpdates = length(dicUpdate);		%number of dictionary updates
+nDecorrAlgs = length(dicDecorr);		%number of decorrelation algorithms
+nCorrLevels = length(minCoherence);		%number of coherence levels
+nInitDicts  = length(initDicts);		%number of initial dictionaries
+
+SMALL.DL(nInitDicts,nCorrLevels,nDecorrAlgs,nDicUpdates) = SMALL_init_DL(toolbox); %create dictionary learning structures
+for iInitDicts=1:nInitDicts
+	for iCorrLevels=1:nCorrLevels
+		for iDecorrAlgs=1:nDecorrAlgs
+			for iDicUpdates=1:nDicUpdates
+				SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).toolbox = toolbox;
+				SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).name = dicUpdate{iDicUpdates};
+				SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).profile = true;
+				SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).param = ...
+					struct( 'data',SMALL.Problem.b,...
+					'Tdata',nActiveAtoms,...
+					'dictsize',dictSize,...
+					'iternum',iterNum,...
+					'memusage','high',...
+					'solver',solver,...
+					'decFcn',dicDecorr{iDecorrAlgs},...
+					'coherence',minCoherence(iCorrLevels),...
+					'initdict',initDicts(iInitDicts));
+				
+				SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates) = ...
+					SMALL_learn(SMALL.Problem,SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates));
+			end
+		end
+	end
+end
+
+%% Evaluate coherence and snr of representation for the various methods
+sr = zeros(size(SMALL.DL));				%signal to noise ratio
+mu = zeros(size(SMALL.DL));				%coherence
+dic(size(SMALL.DL)) = dictionary;		%initialise dictionary objects
+for iInitDict=1:nInitDicts
+	for iCorrLevels=1:nCorrLevels
+		for iDecorrAlgs=1:nDecorrAlgs
+			for iDicUpdates=1:nDicUpdates
+				%Sparse representation
+				SMALL.Problem.A = SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).D;
+				tempSolver = SMALL_solve(SMALL.Problem,solver);
+				%calculate snr
+				sr(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates) = ...
+					snr(SMALL.Problem.b,SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).D*tempSolver.solution);
+				%calculate mu
+				dic(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates) = ...
+					dictionary(SMALL.DL(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).D);
+				mu(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates) = ...
+					dic(iInitDicts,iCorrLevels,iDecorrAlgs,iDicUpdates).coherence;
+			end
+		end
+	end
+end
+
+%% Plot results
+minMu = sqrt((dictSize-blockSize)/(blockSize*(dictSize-1)));	%lowe bound on coherence
+initDictsNames = {'Data','DCT','Gabor'};						
+dicDecorrNames = {'K-SVD','INK-SVD','Grassmannian','New'};
+lineStyles     = {'ks-','kd-','ko-','k*-'};
+for iInitDict=1:nInitDicts
+	figure, hold on, grid on
+	title([initDictsNames{iInitDict} ' Initialisation']);
+	plot([1 1],[0 25],'k-');
+	for iDecorrAlgs=1:nDecorrAlgs
+		plot(mu(iInitDicts,:,iDecorrAlgs,1),sr(iInitDicts,:,iDecorrAlgs,1),...
+			lineStyles{iDecorrAlgs});
+	end
+	plot([minMu minMu],[0 25],'k--')
+	
+	set(gca,'YLim',[0 25],'XLim',[0 1.4]);
+	legend([{'\mu_{max}'},dicDecorrNames,{'\mu_{min}'}]);
+	xlabel('\mu');
+	ylabel('SNR (dB)');
+end