Information Technology Reference
In-Depth Information
Fig. 7.5 Graph of variance
captured by each principal
component in the ORL face
database
M
i
1 e i
i = 1
M
=
where M <
M
N
.
(7.21)
e i
Performance comparison of PCA- and ICA-based feature extraction techniques
in real and complex domain along with different neurons-based classifier is pre-
sented in Table 7.1 . Classifier with complex-valued neurons are fairly better than
with real-valued neurons. Results clearly demonstrate that C RSP and C RSS neuron-
based classifier yield best accuracy with smallest topology. They are far efficient in
comparisonwith any other neurons. The results demonstrating the recognition perfor-
mance in terms of different parameters such as training epochs, number of learning
weights, FRR, FAR, recognition rate are presented in Table 7.1 . The performance
of different ANN architectures are broken down according to the feature extraction
methods. The striking feature of Table 7.1 is that classifier-based on RSP and RSS
neurons in complex domain requires least training epochs and learning parameters
with highest recognition rate. The C ICA-based feature extraction always provides
the best results. Though, difference among different feature extraction methods in the
context of accuracy is not large, but the significant thing is the class distinctiveness
among different subjects. The class distinctiveness is much better in C ICA-based
system in comparison with any other method. Thus, in case of C ICA, the deviation
of square error from threshold is quiet large for class and nonclass faces, whichmakes
it in turn robust to perform more effectively in real life situations where variety of
noises and occlusion may be present in each image. From extensive simulations on
ORL face database, it is imperative to make following inferences:
Class distinctiveness is very poor in case of R PCA and C PCA.
Class distinctiveness with R ICA is slightly better than PCA but it is best with
C ICA.
C ICA-based system is much stricter to unauthorized person.
C RSP neuron-based classifier is outperforming over all other classifiers.
Search WWH ::




Custom Search