Hi Gerhard, Very sorry for the delay. Hopefully now I will have more time to dedicate to lcms stuff.
You are adressing a very interesting question. How to choose the right number of terms, both giving good accurancy and avoiding to fit noise? The answer is: that is no easy :-) Current profiler (and old one too) was using some statistic tools to do that. The key function is cmsxFindOptimumNumOfTerms(). It is doing that by trying each combination and checking the adjusted coefficient of multiple determination. That is a adjusted pearson correlation coefficient, squared. If R2Adj lays between 0.6 and 1, then the correlation is good and I proceed to check standard deviation. The profiler finally picks the number of terms that gives less Std while keeping maximum dE below a huge limit. Sorry if this sounds confusing. You can check the function if wish so, is placed in cmsoutl.c file, and has not changed in new profiler. >I'm also wondering, does your new profiler also take noise statistics >(i.e. standard deviation) of the measurements into account, e.g. for >computing different weights for each patch of the training set, or for >estimating the optimal number of terms to use? Yes, it uses ChiSq value. Some target manufacturers (Kodak) does include it in the sheets. The old measurement tool and the new profiler does also put this parameter in the IT8. You can find some info on that on kodak Q60 ftp: ftp://FTP.KODAK.COM/GASTDS/Q60DATA/ see Chi_sq.pdf and TECHINFO.pdf files. Best regards, Marti. ----- Original Message ----- From: "Gerhard Fuernkranz" <[EMAIL PROTECTED]> To: "Lcms-User" <[EMAIL PROTECTED]> Cc: "Marti Maria" <[EMAIL PROTECTED]> Sent: Tuesday, October 07, 2003 12:40 AM Subject: Re: [Lcms-user] Re: Problems with black point Marti Maria schrieb: >In the other hand, CLUT based can exhibit discontinuities. Thus, a CLUT >based profile can model the weird behavior some devices have. But at a price. >One should be careful on these profiles, because if they are not smooth, they >can fit almost perfectly patches used in training, but be very bad on patches not >used. For example, a profile with dE of about 0.5 can give dE > 10 on new >patches. These profiles are unusable, despite the good dE. > >On more terms in regression, more forced is the gamut to fit outliers, >and less smooth is whole profile. So, using 50 terms is good as far as >20 or less terms gives approximately same result. > Marti, I agree, especially higher order polynomials, if used for curve fitting, tend to match the provided training set very well, but often give a large deviation for data which not in the training set. Additionally, for noisy measurements (and some noise is basically unavoidable), using too many terms rather fits the noise in the data, instead of actually improving the profile. I've also some doubts, whether any reasonable, accurate extrapolation, based on higher order polynomials, beyond the gamut boundary of the IT8 target is possible at all, is it? Is it possible, that LCMS probably also uses more terms than necessary or useful? I'm also wondering, does your new profiler also take noise statistics (i.e. standard deviation) of the measurements into account, e.g. for computing different weights for each patch of the training set, or for estimating the optimal number of terms to use? Regards, Gerhard ------------------------------------------------------- This SF.net email is sponsored by: SF.net Giveback Program. SourceForge.net hosts over 70,000 Open Source Projects. See the people who have HELPED US provide better services: Click here: http://sourceforge.net/supporters.php _______________________________________________ Lcms-user mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/lcms-user
