284 CHAPTER 8. APPLICATIONS
empirical relationships that describe the features of objects. Size, shape, and
other qualitative descriptors provide some level of discrimination between ob-
jects. Selection of these descriptors requires some judgment and testing to see
how well the classifier can be trained to provide a high level of classification.
The descriptors used in this example should by no means be assumed to repre-
sent a unique set. There may be other descriptors, a combination of which could
produce better results. This however is left as an exercise for those attempting
to pursue further studies in this area.
The features used in this example are computed using the following empirical
relationships:
- Form Factor (FF)=(4πA/P^2 )
- Roundness Factor (RF)=4A/πD^2 max
- Aspect Ratio (AR)=mDmax/Dmin
- Compactness Factor (CF)=
p
(4/π)A/DmaxwhereAis the area of an object in pixels,P is the perimeter of an object in
pixels, andDmaxandDminare lengths of the major and minor axes of an ellipse
fitted around an object. Note that the Aspect Ratio is 1.0 for a circular object.
Table 8.19 provides a set of typical features computed for 13 insect classes.
Table 8.19. Typical feature set for various insect classes
Features
Insect Type AR P FF RF CF
Assassin bug 2.4773 109 2.5156 0.2549 0.5049
Big-eyed bug 2.0556 37 4.9074 0.4972 0.7051
Collops beetle 1.7586 51 5.2493 0.5319 0.7293
Cucumber beetle 1.7000 85 5.0970 0.5164 0.7186
Green lacewing adult 4.5000 99 1.8056 0.1829 0.4277
Hippodamia lady beetle adult 1.4375 69 6.4607 0.6546 0.8091
Hippodamia ladybug larva 2.1935 68 3.6090 0.3657 0.6047
Leaf hopper 3.1071 87 2.7873 0.2824 0.5314
Lacewing larva 2.5000 60 3.4566 0.3502 0.5918
Lygus adult 2.4400 61 3.6473 0.3696 0.6079
Nabid adult 2.8000 70 3.1217 0.3163 0.5624
Stinkbug adult 1.4028 101 6.7639 0.6853 0.8278
Three-corned alfalfa hopper 1.6136 71 5.0224 0.5089 0.7134Clustering approach to classifier design Figure 8.25 conceptually illus-
trates clusters of class objects in two-dimensional feature space.