A First Course in FUZZY and NEURAL CONTROL

188 CHAPTER 5. NEURAL NETWORKS FOR CONTROL

Shape Factor: This feature is defined asPerimeter

2

4 π∑Area.

Convex Area: The computation of this feature is somewhat involved. The

best measure for this feature is obtained by constructing a many-sided

polygon around the object and counting the total number of pixels within

the bounding polygon.

Solidity:Thisisdefined asConvex AreaArea.

Convex Perimeter: This feature is computed by bounding an object with a

many-sided irregular polygon and counting the total number of pixels

along the perimeter.

Convexity:Thisisdefined asConvex PerimeterPerimeter.

Maximum and Minimum Diameter: These parameters are obtained by com-

puting the maximum and minimum number of pixels across a many-sided

polygon bounding an object.

Aspect Ratio:Thisisdefined asMaximum DiameterMinimum Diameter.

Compactness:Thisfeatureisdefined as

q

(π^4 )∑Area

Maximum Diameter.

Bounding Box Area: This feature is obtained by determining the area of a

rectangle circumscribing the object with sides of the rectangle parallel to

the image edges.

Extent:Thisisdefined asBounding Box AreaArea.

Edif:ThisisthevariationinExtentwhen an object is rotated by±π 4 radians.

Using the shape descriptors, a set of training patterns and a set of test
patterns were generated for each trash type. For training the neural network, the
targets for bark, stick, leaf, and pepper were arbitrarily chosen as 0.2, 0.4, 0.6,
and 0.8, respectively. Several neural network topologies were examined. In all
cases, the activation function chosen was the unipolar sigmoidal function. The
neural networks were trained to classify trash objects as follows: if the output
is between 0.2±0.1, then the trash object is classified as ìbark.î Similarly,
the object is classified as ìstickî if the output is between 0.4±0.1, ìleafî if
the output is between 0.6±0.1, and ìpepperî if the output is between 0.8±
0.1. Figures 5.14ó5.18 illustrate classification results from the trained neural
networks of various topologies.
In the followingfigures, a two-layer network topology indicated by [10, 1]
implies 10 neurons in thefirst hidden layer and 1 neuron in the output layer.
Similarly, a three-layer neural network topology indicated as [5, 10, 1] implies 5
neurons in thefirst hidden layer, 10 neurons in the second hidden layer, and 1
neuron in the output layer.