Fig. 1a shows some examples of these
misshapen potatoes. The training set consisted of 228 potatoes, of
which 190 were voted as regular tubers and 38 tubers were voted
as irregular ones. A database of images was elaborated by acquiring
an image for each potato by placing them individually inside the
lighting chamber over the conveyor belt. All images were then
saved on the computer to be used as a database to extract the main
features of each potato and to build a linear discriminant analysis
(LDA) model to predict potato regularity. During the learning
mode, the features extracted from the database images of known
shapes were analyzed to predict the regularity of unknown sets
of potatoes. Therefore, another set of potatoes consisting of 182
with 162 regular tubers and 20 misshapen tubers were collected
as a testing set to validate the prediction model in a real-time
application and to evaluate the accuracy of the trained machine vision
system to perform such tasks.