Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 160

Fig. 7.22 Distorted images which are not recognized with any feature extraction-based technique
presented in this topic
7.7 Inferences and Discussion
This chapter is devoted for designing machine learning techniques in complex
domain. The most important contributions are presenting a new class of classifier
“OCON,” which contains the ensemble of higher-order neurons in complex domain
and a new feature extractor “ C ICA” which yielded surprisingly good class distinc-
tiveness among different subjects in image database. The remarkable achievements
in the proposed recognizer (classifier) is its compact structure, improved learning
speed, lesser weights storage and better accuracy in recognition, which contains the
ensemble of single C RSS and C RSP neurons instead of ensemble of neural networks
ie “OCONN.” It is worth mentioning that its overall performance is much better than
recognizer in real domain. The PCA-ICA-based feature extraction algorithms in real
and complex domain are further presented. The capabilities of feature extractor and
classifier are justified through their assessment in typical face recognition system.
The performance of system is tabulated with major performance evaluation metrics.
In order to assess the robustness of different feature extraction techniques to noise,
distortion, and other environmental effects, a comparative assessment of different
techniques over electronically modified images [ 14 ] is carried out. As with any
other classification studies, the class distinctiveness must also be taken into account
for comparing different feature extraction techniques. One major drawback of both
PCA-based feature extraction methods is that distinctiveness among classes during
classification is very poor. They are not able to classify the occluded and blurred
face images considered in experiments. This shows that real and complex PCA are
not suitable for design of recognition system which can work in real environmental
applications. Therefore, the basis vectors obtained by C ICA is superior to R ICA and
PCA in the sense that it provides a representation (feature vector), which yields far
better class distinctiveness, hence is more robust to the effect of noise. It is therefore
observed in this chapter that C ICA outperforms over other techniques for recognition
especially in noisy environment.