pdipm_diff

 PDIPM_DIFF inverse solver for difference data using Primal/Dual interior point method
 img= ab_pdipm( inv_model, data1, data2)
 img        => output image (or vector of images)
 inv_model  => inverse model struct
 data1      => differential data at earlier time
 data2      => differential data at later time

  inv_model.pdipm_diff.norm_data  1 or 2 (DEFAULT 2)
  inv_model.pdipm_diff.norm_image 1 or 2 (DEFAULT 2)
  inv_model.pdipm_diff.beta     (default 1e-6)

 Parameters:
  max_iters =  inv_model.parameters.max_iteration (default 10)
      Max number of iterations before stopping
  min change = inv_model.parameters.min_change   (default 0)
      Min Change in objective fcn (norm(y-Jx)^2 + hp*TV(x)) before stopping
 beta is the parameter that smooths the TV functional

0001 function img=pdipm_diff( inv_model, data1, data2)
0002 % PDIPM_DIFF inverse solver for difference data using Primal/Dual interior point method
0003 % img= ab_pdipm( inv_model, data1, data2)
0004 % img        => output image (or vector of images)
0005 % inv_model  => inverse model struct
0006 % data1      => differential data at earlier time
0007 % data2      => differential data at later time
0008 %
0009 %  inv_model.pdipm_diff.norm_data  1 or 2 (DEFAULT 2)
0010 %  inv_model.pdipm_diff.norm_image 1 or 2 (DEFAULT 2)
0011 %  inv_model.pdipm_diff.beta     (default 1e-6)
0012 %
0013 % Parameters:
0014 %  max_iters =  inv_model.parameters.max_iteration (default 10)
0015 %      Max number of iterations before stopping
0016 %  min change = inv_model.parameters.min_change   (default 0)
0017 %      Min Change in objective fcn (norm(y-Jx)^2 + hp*TV(x)) before stopping
0018 % beta is the parameter that smooths the TV functional
0019 
0020 % (C) 2008 Andrea Borsic. License: GPL version 2 or version 3
0021 % $Id: pdipm_diff.html 2819 2011-09-07 16:43:11Z aadler $
0022 
0023 
0024 pp= process_parameters(inv_model);
0025 
0026 fwd_model= inv_model.fwd_model;
0027 
0028 d=calc_difference_data( data1, data2, fwd_model);
0029 
0030 img_bkgnd=calc_jacobian_bkgnd( inv_model );
0031 J=calc_jacobian( fwd_model, img_bkgnd);
0032 
0033 alpha=calc_hyperparameter( inv_model );
0034 L=calc_R_prior( inv_model );
0035 W= calc_meas_icov( inv_model );
0036 if pp.norm_data==1
0037   W = sqrt(W); % sW is in units of volts
0038 end
0039 
0040 if     pp.norm_data==2 && pp.norm_image==2
0041   x= pdipm_2_2( J,W,alpha*L,d, pp);
0042 elseif pp.norm_data==2 && pp.norm_image==1
0043   x= pdipm_2_1( J,W,alpha*L,d, pp);
0044 elseif pp.norm_data==1 && pp.norm_image==2
0045   x= pdipm_1_2( J,W,alpha*L,d, pp);
0046 elseif pp.norm_data==1 && pp.norm_image==1
0047   x= pdipm_1_1( J,W,alpha*L,d, pp);
0048 end
0049 
0050 % create a data structure to return
0051 img.name = 'pdipm_diff';
0052 img.elem_data = x;
0053 img.fwd_model = fwd_model;
0054 
0055 function s= pdipm_2_2( J,W,L,d, pp);
0056    [s]= initial_values( J, L, pp);
0057 
0058    R = L'*L;
0059    ds= (J'*W*J + R)\(J'*W*(d - J*s) - R*s);
0060    s= s + ds;
0061 
0062 function m= pdipm_1_2( J,W,L,d, pp);
0063    [m,x,jnk,sz]= initial_values( J, L, pp);
0064 
0065    I_M = speye(sz.M, sz.M);
0066    for loop = 1:pp.max_iter
0067       % Define variables
0068       f = J*m - d;             F= spdiag(f);
0069                                X= spdiag(x);
0070       e = sqrt(f.^2 + pp.beta);E= spdiag(e);
0071 
0072       % Define derivatives
0073       dFc_dm = (I_M - X*inv(E)*F)*J;
0074       dFc_dx = -E;
0075       dFf_dm = L'*L;
0076       dFf_dx = J'*W;
0077 
0078       dmdx = -[dFc_dm, dFc_dx; dFf_dm, dFf_dx] \ ...
0079               [ f-E*x; J'*W*x + L'*L*m ];
0080 
0081       dm =             dmdx(      1:sz.N);
0082       dx = x_update(x, dmdx(sz.N+(1:sz.M)));
0083 
0084       m= m + dm; x= x + dx;
0085       loop_display(i)
0086 debug([mean(abs([m,dm])) mean(abs([x,dx]))])
0087       pp = manage_beta(pp);
0088    end
0089 
0090 function m= pdipm_2_1( J,W,L,d, pp);
0091    [m,jnk,y,sz]= initial_values( J, L, pp);
0092 
0093    I_D = speye(sz.D, sz.D);
0094    for loop = 1:pp.max_iter
0095       % Define variables
0096       g = L*m;                 G= spdiag(g);
0097                                Y= spdiag(y);
0098       s = sqrt(g.^2 + pp.beta);S= spdiag(s);
0099 
0100       % Define derivatives
0101       dFf_dm = 2*J'*W*J;
0102       dFf_dy = L';
0103       dFc_dm = (I_D - Y*inv(S)*G)*L;
0104       dFc_dy = -S;
0105 
0106       dmdy = -[dFf_dm, dFf_dy; dFc_dm, dFc_dy] \ ...
0107               [ J'*W*(J*m-d) + L'*y; g-S*y ];
0108 
0109       dm =             dmdy(      1:sz.N );
0110       dy = x_update(y, dmdy(sz.N+(1:sz.D)));
0111 
0112       m= m + dm; y= y + dy;
0113       loop_display(i)
0114 debug([mean(abs([m,dm])), mean(abs([y,dy]))]);
0115       pp = manage_beta(pp);
0116    end
0117 
0118 function m= pdipm_1_1( J,W,L,d, pp);
0119    [m,x,y,sz]= initial_values( J, L, pp);
0120 
0121    I_M = speye(sz.M,sz.M); 
0122    I_D = speye(sz.D,sz.D); 
0123    Z_N = sparse(sz.N,sz.N);
0124    Z_DM= sparse(sz.D,sz.M);
0125    for loop = 1:pp.max_iter
0126       % Define variables
0127       g = L*m;                 G= spdiag(g);
0128       r = sqrt(g.^2 + pp.beta);R= spdiag(r); % S in paper
0129                                Y= spdiag(y);
0130 
0131       f = J*m - d;             F= spdiag(f);
0132       e = sqrt(f.^2 + pp.beta);E= spdiag(e);
0133                                X= spdiag(x);
0134 
0135       % Define derivatives
0136       As1 = Z_N;
0137       As2 = (I_M - X*inv(E)*F) * J;
0138       As3 = (I_D - Y*inv(R)*G) * L;
0139       Ax1 = J'*W;
0140       Ax2 = -E;
0141       Ax3 = Z_DM;
0142       Ay1 = L';
0143       Ay2 = Z_DM';
0144       Ay3 = -R;
0145       B1  = J'*W*x + L'*y;
0146       B2  = f - E*x;
0147       B3  = g - R*y;
0148 
0149       DD = -[As1,Ax1,Ay1; ...
0150              As2,Ax2,Ay2; ...
0151              As3,Ax3,Ay3] \ [B1;B2;B3];
0152 
0153       dm = DD(1:sz.N);
0154       dx = x_update(x, DD(sz.N +        (1:sz.M)) );
0155       dy = x_update(y, DD(sz.N + sz.M + (1:sz.D)) );
0156 
0157       m= m + dm;
0158       x= x + dx;
0159       y= y + dy;
0160       loop_display(i)
0161 debug([mean(abs([m,dm])), mean(abs([x,dx])), mean(abs([y,dy]))]);
0162       pp = manage_beta(pp);
0163    end
0164 
0165 % fix matlab's stupid verbose spdiags function
0166 function sM = spdiag(V)
0167    lV = length(V);
0168    sM = spdiags(V,0,lV,lV);
0169 
0170 function [s,x,y,sz]= initial_values( J, L, pp);
0171    [sz.M,sz.N] = size(J); % M measurements, N parameters
0172    [sz.D  ] = size(L,1); % E edges
0173    y= zeros( sz.D, 1 ); % dual var - start with zeros
0174    s= zeros( sz.N, 1 ); % solution - start with zeros
0175    x= zeros( sz.M, 1 ); % dual var - start with zeros
0176 
0177 % abs(x + dx) must be <= 1
0178 function dx = x_update( x, dx)
0179  % can't have zeros
0180    dx(dx==0) = eps;
0181  % space to limits in direction of dx
0182    sx = sign(dx);
0183    clr = sx - x;
0184    % how much to multiply by to get to limits
0185    fac = clr./dx;
0186    % choose min amount to get to limits
0187    dx = dx*min(abs(fac));
0188 %  dx = dx*0.9;
0189 
0190 function debug(vals)
0191 %  disp(vals)
0192 
0193 function pp = manage_beta(pp);
0194    pp.beta = pp.beta * pp.beta_reduce;
0195    if pp.beta < pp.beta_minimum;
0196       pp.beta = pp.beta_minimum;
0197    end
0198 
0199 function pp= process_parameters(imdl);
0200    try    pp.max_iter = imdl.parameters.max_iterations;
0201    catch  pp.max_iter = 10;
0202    end
0203 
0204    try    pp.min_change = imdl.parameters.min_change;
0205    catch  pp.min_change = 0;
0206    end
0207 
0208    try    pp.beta = imdl.pdipm_diff.beta; 
0209    catch  pp.beta = 1e-6;
0210    end
0211 
0212    pp.beta_reduce = 0.2;
0213    pp.beta_minimum= 1e-16;
0214 
0215    try    pp.norm_data = imdl.pdipm_diff.norm_data;
0216    catch  pp.norm_data = 2;
0217    end
0218 
0219    try    pp.norm_image = imdl.pdipm_diff.norm_image;
0220    catch  pp.norm_image = 2;
0221    end
0222 
0223 function loop_display(i)
0224    fprintf('+');