> -------- Original-Nachricht -------- > Datum: Wed, 24 May 2006 11:02:51 +1000 > Von: Graeme Gill <graeme@xxxxxxxxxxxxx> > > > Like I said, I used 39 test patches from an Argyll 39-patch chart, > > with eight copies of each patch. > > OK, seems pretty reasonable then. Except probably, that a sample size of 8 per color is rather very small to estimate the population variance by the sample variance reliably. I have done basicallly the same experiment some time ago (also with about 50 patches and 10 replications each) in order to get an estimate of the reproducibility of my laser printer. Basically there is a trend in evidence that some colors are reproduced with a lower and others with a larger variance, however one must be very careful judging this, since for such a small sample size, a difference between say 1.8dE and 2dE is statistically not really significant; i.e. the color with 1.8dE sample standard deviation and the other color with the 2dE sample standard deviation could very likely also be just two different samples from the same population, or the color with the 1.8 dE sample standard deviation could even come from a population with a larger population standard deviation then the color with the 2dE sample standard deviation. I'm in fact wondering whether it might be worth to attempt to fit a very smooth RGB/CMYK -> variance RSPL mapping to this data, and to use this RSPL during profile generation to obtain an individual weight for each training set data point (i.e. weight[i]=1/variance[i], probably normalized such that the sum of the weights is 1.0 in order that the per data point weights do not interfere to much with the existing avgdev -> raw smoothness mapping). Regards, Gerhard > The real overhead is in making > several separate prints to exercise that source of error. -- Gerhard Fuernkranz nospam456@xxxxxx Bis zu 70% Ihrer Onlinekosten sparen: GMX SmartSurfer! Kostenlos downloaden: http://www.gmx.net/de/go/smartsurfer