186 CHAPTER 5. NEURAL NETWORKS FOR CONTROL
perform trash identification. We explain here how a backpropagation neural net-
work was trained to identify four different trash types in cotton, namely, bark,
stick, leaf, and pepper [65]. The training data used to train the neural network
consisted of a large number of samples for which image processing techniques
were used to determine the features of each trash type. Different neural network
topologies were trained and tested to evaluate their classification performance.
Results from two neural network topologies ñ one with a single hidden layer
and an output layer and the other with two hidden layers and an output layer
ñ are presented here for comparative purposes.
The fundamental objective in object recognition is to utilize a minimum
number of features that can be used not only to identify a specific object, but
also to distinguish between objects. To accomplish this objective, we need to
obtain the right combination of distinguishing features. Extracting features
of objects that do not have any mathematical representation is based upon
empirical relationships that provide some measure of the object shape. Objects
such as ìbark,î ìstick,î ìleaf,î and ìpepperî have characteristics that must
be obtained from image analysis of such objects. Figures 5.10ó5.13 illustrate
samples of trash objects and their corresponding computer-segmented images
that were used in obtaining features. Numbers alongside the segmented objects
are object identifiers that are used to tag the computed features to that object.
Figure 5.10. Bark sample and segmented imageFigure 5.11. Stick sample and segmented image