How can i use gabor filter for face recognition

Hi I am writing a matlab code to recognize the face using Gabor filter. I have created a Gabor filter bank using central frequency of 0.7 and Theta = 0, pi/3,2*pi/3 and pi. These gabor filters are then convolued on the image database ,After this PCA was used on filter image using matlab command of princomp, the eigen vectors so obtained are than used to project the images into the subspace to generate the feature vectors . Once the feautre vector is obtained than the simillarity input image with the training images is calulted is done using the Euclidian distance ,However i found that it did not give me correct result i.e the images were not correctly recognized. My question is should i also used the input image (i.e. one which need to be recognized from the training set of images) by first filter it throung the same bank of gabor filter which i used to create the feature space. any suggestion???

Answer 1

[A]

approach is right but not optimum. you may try for more gabor filter by making set of different frequencies and theta, lets 5X8. then make a combined vector and pCA and euclidean distance. I suggest you to have gabor filter coding accurate as it is core part of your recognition work.

[nadeem]

Thnx, but I have observed as increasing the gabor frequency actually produced filter image with mor edges and that actually form a very blur and deform the images very much. Btw what would be ur suggestion for the range of frequency to be used. In my Programe i am actually using the Lambada the formula I am using is :

gb=exp(-0.5*(x_theta.^2/sigma_x^2+y_theta.^2/sigma_y^2)).*cos(2*pi/lambda*x_theta+psi);

where

x_theta=x*cos(theta)+y*sin(theta);

y_theta=-x*sin(theta)+y*cos(theta);

Any comments on it.

[jeba shanthi]

please explain breif...

Answer 2

[Peter Manley-Cooke]

You will also have to convert your image to be recognised to the same feature space as your samples.

[nadeem]

thnx Peter, Following ur approach do u mean that i actually should apply the whole bank of Gabr filters that i used during training to the test image as well. If Yes than this mean i would have another feature space of the test image say for example 5X8 and than I should find the Euclian distance .........

Answer 3

[jeba shanthi]

can i give gabor filter out for recognition input? am i correct or not?