Databases Reference
In-Depth Information
(a) Mean pixel value.
(b) Mean pixel value,
µ
vs.
standard deviation, σ.
FIGURE 7.1: Data feature representation for a set of images.
algorithm must determine and continue to adjust synaptic weights during the
recognition process [87]. To ensure the e
ciency of the recognition scheme, it
is imperative that the number of features be kept to a minimum. Neverthe-
less, selecting features for recognition is a complex process that needs to be
performed objectively.
To address the curse of dimensionality, current approaches extend the recog-
nition process by introducing a feature selection mechanism to select the fea-
tures that best represent the entire data set. However, dimensionality reduc-
tion adds to the complexity of the recognition processes and requires the use of
costly algorithms, such as the principal component analysis (PCA). Further-
more, erroneous feature selection can affect the accuracy of the recognition
scheme. A simple recognition scheme that is capable of analyzing more than
one feature and does not use a feature selection mechanism to determine the
best features for data representation is needed.
7.2 Distributed Multi-Feature Recognition
The scalability of commonly used pattern recognition techniques involving
multiple features usually deteriorates as the number of training and testing
data sets increases. In this chapter, we will look at multi-feature recognition
by including the distributiveness that occurs in natural schemes. The DHGN
algorithm has been modeled with a fully distributed approach for recogni-
tion using multiple data features. The following subsections will describe the
distributed DHGN scheme for multi-feature recognition.
The DHGN multi-feature scheme conducts distributed pattern recogni-
tion using features obtained from pattern data through a feature extraction
method. It provides a scalable approach; the number of features required for
