diff DL/Majorization Minimization DL/dict_update_REG_cn.m @ 155:b14209313ba4 ivand_dev

Integration of Majorization Minimisation Dictionary Learning

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/DL/Majorization Minimization DL/dict_update_REG_cn.m	Mon Aug 22 11:46:35 2011 +0100
@@ -0,0 +1,48 @@
+function [Phiout,unhatnz] = dict_update_REG_cn(Phi,x,unhat,maxIT,eps,cvset)
+%% Regularized Dictionary Learning with the constraint on the column norms %%%%%
+% Phi = Normalized Initial Dictionary
+% x = Signal(x). This can be a vector or a matrix
+% unhat = Initial guess for the coefficients
+% to = 1/(step size) . It is larger than spectral norm of coefficient matrix x
+% eps = Stopping criterion for iterative softthresholding and MM dictionary update
+% cvset = Dictionary constraint. 0 = Non convex ||d|| = 1, 1 = Convex ||d||<=1
+% Phiout = Updated Dictionary
+% unhatnz Updated Coefficients (the same as input in this version)
+
+%%
+B = Phi;
+K = zeros(size(Phi,1),size(Phi,2));
+c = .1 + svds(unhat,1)^2; %.1
+c3 = (1/c)*eye(size(B,2));
+c1 = x*unhat'*c3;
+c2 = (c*eye(size(B,2))-unhat*unhat')*c3;
+
+%%
+      
+for i=1:maxIT
+%     if i>1
+%         B = K;
+%     end
+
+    K = c1 + B*c2;
+    
+    if cvset == 1,
+        K = K*diag(min(sum(K.^2).^(-.5),1)); % with convex constraint set
+    else
+        % Mehrdad original - 
+        % K = K*diag(sum(K.^2).^(-.5)); % with fixed-norm constraint set
+        K = normc(K);
+    end
+       
+    if (sum(sum((B-K).^2)) < eps)
+        break;
+    end
+    
+    B = K;
+end
+%% depleted atoms cancellation %%%
+[Y,I] = sort(sum(K.^2),'descend');
+RR = sum(Y>=.01);
+Phiout = K(:,I(1:RR));
+unhatnz = unhat(I(1:RR),:);
+end
\ No newline at end of file