部署PCA的gabor滤波器组中的功能减少

Question

Using Matlab, I'm working with Gabor filter bank, with different orientations and scales, I got a huge no.of features by the no.of used filters. with the total number of training data, I want to deploy the PCA to reduce the features number, but I don't know how to begin and which function to use, plz help.

Answer 1

If you want to use PCA in order to reduce the number of features Matlab makes everything available to you

A little bit more basic version:

%% some data points for 2 features
X = [1,1,1, 2,2,2, 3,3,3 ; 1,2,3, 2,3,4, 3,4,5];
numX = size(X,2);

%% PCA
covX = zeros(2,2); % sample covariance
meanX = mean(X,2); % sample mean
for k = 1:numX
    covX = covX + (X(:,k) - meanX) * (X(:,k) - meanX)';
end
covX = covX / (numX - 1);

% essential part of PCA:
[V,D] = eig(covX);

Now feature reduction by PCA usually means to interpret the directions of highest variance (corresponding to diagonal entries in D ) as most discriminative ones. So you can use just a few columns of V to reweight your features to a smaller subset of largest variation.

Maybe a little bit off your question but quite important in my view: Deploying PCA is not always the best choice to reduce the number of features since the directions of largest variance are not always the most discriminating directions! Furthermore, you still have to compute all your features and just reduce the number for classification (which hopefully saves you from overfitting problems). So maybe you want to consider other feature selection methods as well.