Spectral Transformations - Biomedical Signal Analysis: Contemporary Methods and Applications - page 29

2

Spectral Transformations

Pattern recognition tasks require the conversion of biosignals in features
describing the collected sensor data in a compact form. Ideally, this
should pertain only to relevant information. Feature extraction is an im-
portant technique in pattern recognition by determining descriptors for
reducing dimensionality of pattern representation. A lower-dimensional
representation of a signal is a feature . It plays a key role in determining
the discriminating properties of signal classes. The choice of features, or
measurements, has an important influence on (1) accuracy of classifica-
tion, (2) time needed for classification, (3) number of examples needed
for learning, and (4) cost of performing classification.
A carefully selected feature should remain unchanged if there are
variations within a signal class, and it should reveal important differences
when discriminating between patterns of different signal classes. In other
words, patterns are described with as little loss as possible of pertinent
information.
There are four known categories in the literature for extracting
features [54]:

1. Nontransformed structural characteristics: moments, power, amplitude
information, energy, etc.
2. Transformed signal characteristics: frequency and amplitude spectra,
subspace transformation methods, etc.
3. Structural descriptions: formal languages and their grammars, parsing
techniques, and string matching techniques
4. Graph descriptors: attributed graphs, relational graphs, and semantic
networks

Transformed signal characteristics form the most relevant category
for biosignal processing and feature extraction. The basic idea employed
in transformed signal characteristics is to find such transform-based
features with a high information density of the original input and a
low redundancy. To understand this aspect better, let us consider a
radiographic image. The pixels (input samples) at the various positions
have a large degree of correlation. Gray values only introduce redundant
information for the subsequent classification. For example, by using
the wavelet transform we obtain a feature set based on the wavelet