Biomedical Engineering Reference
In-Depth Information
similar or equivalent in some way, samples are classified programmatically into empirically
established groups, based on groups or clusters in the unlabeled collection of samples. That
is, simple pattern recognition is assumption-driven, in that a hypothesis is developed and
tested against the data. In pattern discovery, the extracted data serve as the seed of a new
hypothesis. Clustering techniques are used to group samples that are more similar to each
other than to other groups, and that have a low internal cluster variability or scatter.
Measurement. The measurement phase of the pattern-recognition and discovery process
involves converting the original pattern into a representation that can be easily manipulated
programmatically. For example, a 3D vector image of a protein might be represented as a
series of 2D matrices. Similarly, a nucleotide sequence may be represented by a series of
integers (for example, A = 1, T = 2, C = 3, and G = 4), depending on the underlying
technology used to perform the pattern-matching operation.
l
Processing. After the measurement process, the data are processed to remove noise and
prepare for feature extraction. Processing typically involves executing a variety of error
checking and correction routines, as well as specialized processes that depend on the nature
of the data. For example, images may undergo edge enhancement and transformation to
correct for size and orientation variations (normalization) in order to facilitate feature
extraction.
l
Feature Extraction. Feature extraction involves searching for global and local features in
the data that are defined as relevant to pattern matching during feature selection. Clustering
techniques, in which similar data are grouped together, often form the basis of feature
extraction.
l
Classification and Discovery. In the classification phase of pattern recognition and
discovery, data are classified based on measurements of similarity with other patterns. These
measurements of similarity are commonly based on either a statistical or a structural
approach. In the statistical approach, exemplar patterns are represented by points in a
multidimensional space that is partitioned into regions associated with a classification. In the
structural approach, the structures of the exemplar patterns are explicitly defined. In either
case, the similarity of the data to be classified is compared with the exemplar data to assess
closeness of association.
l
Labeling. The pattern-recognition process ends when a label is assigned to the data, based
on its membership in a class.
l
As illustrated in Figure 7-5 , the pattern-recognition process isn't unidirectional, but is iterative to the
extent that failures at the classification and feature-extraction stages can be corrected by
reevaluating the preceding phase. For example, if the feature-extraction phase fails to identify
relevant data, then the processing of the original image may need to be modified by removing
extraneous data from consideration and by taking other, more relevant data, into consideration.
Feature extraction and classification and discovery, which represent the core of the pattern-
recognition and discovery process, are performed by using some combination of classification,
regression, segmentation, link analysis, and deviation detection methods, depending on the nature of
the data. Similarly, these methods are supported by a variety of technologies and approaches,
collectively referred to as machine learning, as described here.
Search WWH ::




Custom Search