A First Course in FUZZY and NEURAL CONTROL

276 CHAPTER 8. APPLICATIONS

pepper objects at high accuracies, the computation of the classification rates
might be skewed as a classification criterion. However, it should be mentioned
thatbasedontheclassification results of the training data, it is evident that it
is possible to obtain superior classification rates for objects that prescribe the
typical shapes that define the trash types. Also as mentioned previously, most
of the misclassification is due to segmentation defects. Research efforts with
segmentation techniques and acquisition of cotton images need to be addressed
in the future.

Table 8.11. Classification results of trash objects

infive test samples using ANFIS: 232-partition

Total number Actual Classified trash type

of objects trash type Bark Stick Leaf Pepper

10 Bark 3 4 3 0

4Stick0103

16 Leaf 1 2 6 7

215 Pepper 0 0 0 215

In order to evaluate the performance of the ANFIS, different membership
partitions were tested. The classification accuracy of the trash types with the
use of triangular and trapezoidal membership functions was also investigated.
The best classification results were obtained with the membership partition as
shown in Figures 8.18ó8.20 for the three features.

Initial MF's on Area

0

0.2

0.4

0.6

0.8

1

1.2

0 10002000300040005000

Area

Final MF's on Area

0

0.2

0.4

0.6

0.8

1

1.2

0 1000 2000 3000 4000 5000

Area

Figure 8.18. Initial andfinal membership functions onarea: 222-partition

Initial MF's on Solidity

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1

Solidity

Final MF's on Solidity

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1

Solidity

Figure 8.19. Initial andfinal membership functions onsolidity: 222-partition