parzen_map.m

模式识别工具包

%PARZEN_MAP Map a dataset on a Parzen densities based classifier
% 
% 	F = parzen_map(A,W)
% 
% Maps the dataset A by the Parzen density based classfier W. F*sigm 
% are the posterior probabilities. W should be trained by a 
% classifier like parzenc. This routine is called automatically to 
% solve A*W if W is trained by parzenc.
% 
% The global PRMEMORY is read for the maximum size of the internally 
% declared matrix, default inf.
% 
% See also mappings, datasets, parzenc, testp

% Copyright: R.P.W. Duin, [e-mail]
% Faculty of Applied Physics, Delft University of Technology
% P.O. Box 5046, 2600 GA Delft, The Netherlands

function F = parzen_map(T,W)
if nargin < 2, W = parzenc(T); end
[a,classlist,type,k,c,v,h] = mapping(W);
if ~strcmp(type,'parzen_map') 
	error('Wrong type of classifier')
end
[nlab,lablist,m,k,c,p] = dataset(a);
p = p(:)';
h = h(:)';
[mt,kt] = size(T);
if kt ~= k, error('Wrong feature size'); end

if length(h) == 1, h = h * ones(1,c); end
if length(h) ~= c
	error('Wrong number of smoothing parameters')
end
maxa = max(max(abs(a)));
a = a/maxa;
T = T/maxa;
h = h/maxa;
if isfeatim(T)
	F = datgauss(T,h);
end
alf=sqrt(2*pi)^k;
[num,n] = prmem(mt,m);
F = ones(mt,c);
for j = 0:num-1
	if j == num-1
		nn = mt - num*n + n;
	else
		nn = n;
	end
	range = [j*n+1:j*n+nn];
	D = +distm(a,T(range,:));
	for i=1:c
		I = find(nlab == i);
		if length(I) > 0
			F(range,i) = mean(exp(-D(I,:)*0.5./(h(i).^2)),1)';
		end
	end
end
%F = F.*repmat(p./(alf.*h.^k),mt,1);
if max(h) ~= min(h)	% avoid this when possible (problems with large k)
	F = F.*repmat(p./(h.^k),mt,1);
else
	F = F.*repmat(p,mt,1);
end
F = F + realmin;
F = F ./ (sum(F')'*ones(1,c));
F = invsig(F);
[nlab,lablist,m,k,c,p] = dataset(T);
F = dataset(F,getlab(T),classlist,p,lablist);
return