Am 16.11.2011 15:07, schrieb Alexey Gribunin:
By the way, as I understand, number of principal components should be equal or less than 36, i.e. maximum spectral bands. Is this correct?
Yes.
Why did you use 15?
I have selected this value empirically for my data set. The reflectance spectra of all the 1786 patches in my data set (which was btw. from an inkjet print, measured with Spectrolino) can be represented as linear combinations of 15 basis vectors with an error of less than 0.01 dE average, and 0.1 dE max, which is good enough and certainly below the repeatability of the measurements. Thus more than 15 PCs would just fit more of the noise in the measurements, and not really improving anything (likely I should have used even less than 15 for this data set). Regards, Gerhard