Help MACE filter design [Matlab]

Prev: line through a cylinder
Next: making executable file from Guide GUI

From: tarik bahraoui on 8 Mar 2010 14:03

"hulijo" <u44586(a)uwe> wrote in message <867359da739e0(a)uwe>...
> Hy yonnas,
> It seems that I am having a similar problem trying to implement the MACE
> filter. I am quite sure my code is wright, but the results I get when using
> MACE correlation filters for recognition are nothing like those reported in
> the literature. I achieve lets say a equal error rate of 8% in my experiments
> which, however, is comparable only to PCA not even to LDA or other
> discriminative methods.
>
> I suspect two things, at least what face recognition is concerned. First,
> most papers use a database were pleanty of images per client are available
> for training, while I use only three. And second they perform identitfication
> experiments, while I do verification. Anyhow, my code for constructing MACE
> filters is a bit different from yours as I use training images of all
> subjects to construct the filters.
>
> The code:
>
> %read files in the list called "ucn1" and do zero-mean-unit-variance
> normalization
> vector_size=10000;
> number_of_training_images = 600;
> Y=[];
> for i=1:number_of_training_images
> X = double(imread(ucn1(i,:)));
> x = reshape(X,[vector_size,1]);
> means = mean(x);
> vari = sqrt((sum((x-means).^2))/vector_size);
> x = (x-means)/vari;
> Y=[Y,x];
> end
> Y=Y';
>
> %transfrom to frequenca domain
> im_hight = 100;
> im_width = 100;
> for i=1:number_of_training_images
> X = reshape(Y(i,:),[im_hight,im_width]);
> X_f = fft2(X);
> X_f = reshape(X_f,[vector_size,1]);
> Y_f(i,:)=X_f';
> end
> Y_f = Y_f';
>
> % MACE filter with all training images, result in "h"
> counter=1;
> X2 = Y_f;
> D=mean((abs(X2)),2));
> faktor = repmat(1./D,1,number_of_training_images).*X2*inv((X2)'*(repmat(1./D,
> 1,number_of_training_images).*X2));
> for i=1:3:number_of_training_images %I have three images per
> subject
> c=zeros(number_of_training_images,1);
> c(i:i+2,1)=1;
> h(:,counter) = (faktor*c);
> counter=counter+1;
> end
>
> %h now contains all the filters
>
>
>
> %for a test image
> X = double(imread(ucn1(1,:)));
> x = reshape(X,[vector_size,1]);
> means = mean(x);
> vari = sqrt((sum((x-means).^2))/vector_size);
> x = (x-means)/vari;
> X=reshape(x,[im_hight,im_width]);
> out = real( ifftshift((ifft2(((fft2(X)).*(reshape(conj(h(:,1)),[im_hight,
> im_width])))))));
>
>
>
> I hope it the code will help you!
>
> Regards,
>
> hulijo
>
>
> yonnas syams wrote:
> >Hi everyone,
> >I need help in Minimum Average Correlation Energy Filter
> >(MACE)filter design.I am trying to make it according to
> >http://s.i-techonline.com/Book/Face-Recognition/ISBN978-3-902613-03-5-fr26.pdf
> >
> >I think my program doesn't work well.When I apply it to the
> >training image used, it's true that there is peak at the
> >correlation output (positive or negative) peak.But when I
> >apply the testing image(from the same class/person), there
> >is no peak at all, and i think there's should be a peak.
> >One more question,how to normalized the correlation output
> >so in will be in range -1 to 1.
> >Any help will be greatly appreciated.
> >
> >Here is my codes:
> >%Filter Design
> >% 3 iris input training , N=3
> >ans1=imread('001_1_1.bmp');
> >ans2=imread('001_1_2.bmp');
> >ans3=imread('001_1_3.bmp');
> >%change to domain frequency x[d1 d2]
> >x1=fft2(im2double((ans1.hasil));
> >x2=fft2(im2double((ans2.hasil));
> >x3=fft2(im2double((ans3.hasil));
> >%change to vector kolom X[d 1]
> >X1 =x1(:);
> >X2 =x2(:);
> >X3 =x3(:);
> >%matrix X[d N]
> >X=[X1 X2 X3];
> >%the output i wanted, 1 for authentic image
> >u=[1 1 1]';
> >[d N]=size(X);
> >%compute D, in the paper D is diagonal matrix, but not here
> >because the result is the same
> >D1 =X1 .* X1;
> >D2 =X2 .* X2;
> >D3 =X3 .* X3;
> >D =(1/N*d) * (D1+D2+D3) ;
> >%make inverse D
> >for i=1:d
> > D_inv(i,1)=1/D(i,1);
> >end
> >%a= inverse(ctranspose X[N d] * inverse D[d 1] * X[d N])
> >%temp[d N]=inv D[d 1] * X[d N]
> >for i=1:d
> > for j=1:N
> > temp(i,j) = D_inv(i,1) * X(i,j);
> > end
> >end
> >%a[N N]=inv(Xt[N d] * temp[d N]);
> >a =inv(ctranspose(X) * temp);
> >%h[d 1]=(inverse D[d 1] * X[d N])[d N] * a[N N] * u[N 1];
> >h =temp * a * u;
> >%mace filter in domain frequency
> >%change to the original size [d1 d2]
> >[d1 d2]=size(x1);
> >Hmace=reshape(h,d1,d2);
> >save Hmace;
> >
> >%Aplly to Image
> >iris=imread('001_2_1.bmp');
> >load Hmace;
> >x=fft2(im2double(iris));
> >c=ifft2(x .* Hmace);
> >figure, mesh(c)
> >
> >Note : My english is not so good, i am sorry ^_^
>

From: tarik bahraoui on 8 Mar 2010 14:05

|
Pages: 1
Prev: line through a cylinder
Next: making executable file from Guide GUI