learn_snapshots.m

Standard model object recognition matlab code

function [f1,f2] = learn_snapshots (X,patch_siz,num_patch,num_afferents)

% FUNCTION [f1,f2] = learn_snapshots (X,patch_siz,num_patch,num_afferents)
%
% Take num_patch snapshots of object X, which is a 3-D matrix.  These
% snapshots have the dimensions of patch_siz (in the first two
% dimensions of X) and the num_afferents number of afferents.

global ON_CON;
[maxY, maxX, num_feat] = size(X);

f1 = [];
f2 = [];

% Get patches
PatchSiz = max(patch_siz);

if ((maxX-PatchSiz) < 0 | (maxY-PatchSiz) < 0)
  disp('Image is too small for learning patches: Skip image.');
else
  tot_pos  = (maxX-PatchSiz+1)*(maxY-PatchSiz+1);

  for k = 1:length(patch_siz)
    rand_pos = randperm(tot_pos);

    if (num_patch > tot_pos)
      disp(['Max number of patches (' num2str(tot_pos) ') exceeded']);
      nPatch = tot_pos;
    else
      nPatch = num_patch;
    end

    for l = 1:nPatch
      [y, x]    = ind2sub([maxY-PatchSiz+1 maxX-PatchSiz+1], rand_pos(l));
      feat      = X(y:y+PatchSiz-1, x:x+PatchSiz-1, :);
      tot_pos2  = patch_siz(k)*patch_siz(k)*num_feat;
      rand_pos2 = randperm(tot_pos2);
      mask      = zeros([size(feat)]);
      norm      = 0;

      for p = 1:num_afferents
	indx      = [patch_siz(k) patch_siz(k) num_feat];
	[y2,x2,f] = ind2sub(indx, rand_pos2(p));

	% Since we may want to keep zero weight, we have a global
	% variable ON_CON that specifies the "on connections".
	% Any weights that are greater than ON_CON will be known
	% to be the valid weight.  (Although f2 would also act as
	% ON_CON identifier...)
	mask(y2,x2,f) = ON_CON;
	norm          = norm + feat(y2,x2,f);
      end

      if norm > 0.01  % if the patch is not blank.
	% hence the total number of features may be smaller than nPatch.
	f1 = cat(4, f1, feat+mask);
	f2 = cat(4, f2, mask == ON_CON);
      end
      
    end
  end
end