In a domain where it is costly in terms of time resources to
gather large datasets of samples, the importance of lowering
the dimensionality of the learning problem becomes even
more acute. The domain of fruit sorting is one such area,
since each image sample must be carefully inspected and
correctly labeled with the appropriate class. The learning
problem for fruit classification in this case lends itself well
to this form of sub-tasking which can be reformulated into a
hierarchical decomposition, where the outputs of one subproblem
become the inputs to another. In this instance,
features relevant to blemish classification are extracted and
used only for the learning of the blemish classifiers, whose
output becomes the new input feature for the induction of
the global fruit grading classifier.