A machine vision system was developed to inspect cut roses and sort in
to quality categories similar to those used by human inspectors. Image
processing techniques were developed to find the base of the stem, th
e top of the bud, visible portions of the stem, and the projected area
of the bud. Quantitative features were identified to analyze rose qua
lity, including stem length, stem diameter, stem straightness, bud mat
urity, and bud color. Bayes decision theory was used to develop a clas
sifier for straightness and maturity. Straightness was also classified
by a neural network. Experimental tests were run on commercially prod
uced yellow and white roses ('Yellow Waves' and 'White Mystery'). The
machine vision system measured stem length with an average absolute er
ror of 7 mm (2.2% relative error) and stem diameter with an average ab
solute error of 0.6 mm (16% relative error). Straightness classificati
on errors for the yellow cultivar were 17% with the Bayes classifier a
nd 18% with the neural network. All errors with the Bayes classifier w
ere due to misclassification of crooked roses (30%), whereas errors wi
th the neural network were due to misclassification of both crooked an
d straight roses (15 and 22%, respectively). Problems with identificat
ion of the stem segments due to foliage caused most of these errors. M
aturity classification errors with the Bayes classifier were 15% for t
he yellow cultivar and 21% for the white cultivar. Although it was rel
atively easy to separate tight (immature) buds from more open (mature)
buds due to the large difference in features between classes, it was
more difficult to distinguish slight degrees of openness. The machine
vision system was able to accurately separate the cultivars by color u
sing the b chromaticity value of the bud.