Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 159

As with any other classification studies, the class distinctiveness must be taken
into account for comparing different feature extraction techniques. In order to further
access the robustness of different feature extraction techniques to noise, distortion,
and other environmental effects, I carried out a comparative assessment of different
techniques over electronically modified [ 14 ] coloured images. Figure 7.20 presents a
set of images in which low distortion is introduced. These images are passed to pre-
viously trained C RSP classifiers with different feature extraction techniques. C ICA
and R ICA -based classifiers are able to correctly classify them while any PCA-based
classifiers are not able to recognize them. Another set of images in Fig. 7.21 have
comparativelymore distortions and variations; these images can not be recognized by
any PCA and R ICA -based recognizer, while C ICA -based recognition system is able
to correctly identify them. It is due to better class distinctiveness yielded by C ICA .
Though, there are limitations in C ICA -based technique as there are always in every
methods. Figure 7.22 presents some images with very high degree of distortions,
where C ICA -based system also failed to perform proper recognition. Therefore, the
basis vectors obtained by C ICA is superior to R ICA and PCA in the sense that it pro-
vides a representation (feature vector), which is more robust to the effect of noise. It
is therefore observed in all experiments that C ICA outperforms over other popular
techniques, PCA and R ICA , for recognition specially in noisy environment.
Fig. 7.20 Distorted images which are not recognized with any PCA -based recognition system. But,
recognized by R ICA and C ICA -based recognition system
Fig. 7.21 Distorted images which are recognized with only C ICA -based recognizer