I wrote this in reply to a posting on the ASQ forum. I am a
retiree of 3M Company and now have my own company that provides software for developing
acceptance sampling plans.
I have done a lot of work using L,a,b and x,y,z,x' color
space. I prefer acceptance criteria that are not only within perception-related
requirement like deltaE<=1, but that is additionally based on product/measurement
I think the acceptance criterion of DeltaE<=1 may be a
wider interval than it has to be if you would base it on capability. In my experience,
DeltaE has had a within-lot standard deviation close to 0.065 deltaE units. If this is the
case with your product/measurement system, then you can measure much more precisely than
you can see! (Using DeltaE=1 as a definition of visible difference.)
Acceptable Mean of Delta-E
The acceptance criterion would be the acceptance limit of a
sampling plan for the mean of deltaE having an acceptable mean (AQL) of 0.065*3=.195 deltaE units. (Of course you would use
a value for the within-lot standard deviation derived from your specific materials and
Rejectable Mean and Sampling Risks
Then you would choose a rejectable mean (RQL) a little larger than AQL=.195, maybe 1.5 standard
deviations above AQL, chosen such that the sample size is fairly small. I would set the
sampling risks to alpha=0.05 and beta=0.05.
The same method can be applied to L, a, and b individually,
which works well if you use the separate within-lot standard deviations of L, a, and b.
Sequential Sampling Plan and Range
I have found sequential
sampling plans for the mean to be very efficient and practical with color
applications, with a minimum n of 2 so that outliers can be detected with a range test.
If you are interested in seeing information on software that have developed for designing this type of plan,
and further explanation of the methodology, just send me your fax number by email and I
will fax you a 4 page description. Or, send a mailing
In summary, I would use acceptance criteria related to
process/measurement capability rather than perception-related numbers like ± 1.0 or ±
0.5. The suggested methodology is statistically based.