How to use PCA as feature descriptor for images like FFT, GLCM etc???Please help

I want to use PCA for texture image feature extraction. Can I use this like FFT, GLCM does like mean_fft, var_fft?? Can I use like this way..

Use this mean,var as a feature vector or input for classifier to classify the category of images..

Can i use this?Is it feasible theoretically..??

Plz help with sample code..

Answer 1

[Image Analyst]

Just treat the PC image like any other image, for example

meanOfPC1 = mean(PC1(:)); % Get mean of the first PC image.
sdOfPC1 = std(PC1(:)); % Get standard deviation of the first PC image.

[begginer01]

Thank you for the ans..I have few confusion..How to calculate PC image??.. we can use coeff and then transform it to get PC image? Am I right?

After the mean and sd calculation of pc image,,can we use this values as a feature input for classification purpose??Is this a feasible method for classification??

Please help with sample code..

[Image Analyst]

See attached demo.

[begginer01]

Thanks for the help. After getting this PCA image..can i use this image for feature extraction and find mean and sd of this image becz theory says PCA mainly uses for dimension reduction..Plz clear my doubt..can i apply this method for classification?

[Image Analyst]

You can use it both ways. If you get the PCA images, then you can use just one, say the first one which will be most dominant, for feature extraction, and then discard/ignore the others if you want (dimension reduction is performed by essentially ignoring worthless, insignificant, unnecessary PCs).

For example, say you have pink objects on a gray background. The first PC will be the overall brightness image, like it you had done rgb2gray(). You might be able to threshold this PC1 image to detect the objects shape and size. The other PCs are related to the color of the objects but maybe the fact that it's pink is not needed because you can find it simply by the brightness in PC1. So you can ignore PC2 and PC3 because the color information does not help you.