Improving the Recognition Performance of Moment Features by Selection - Feature Selection for Data and Pattern Recognition - page 297

The successful operation of the pattern recognition procedure is mainly based on
the representationmethod of the real patterns in a form suitable to be manipulated and
managed by the recognition module (classifier). In the case of image based systems
the content of an image pattern has to be transformed in a compact numerical format
(or other) by applying a feature extraction method (FEM). The role of a FEM is
twofold; it performs a dimensionality reduction from the space of image pixels to a
small set of numbers and it captures the discriminative characteristics of the patterns
in order to distinguish them.
A popular feature extraction method for the case of image patterns is the method
of moments. Image moments have proved to be efficient descriptors of the images'
content, with many applications in pattern recognition [ 2 , 22 , 25 , 36 ], computer
vision [ 12 , 23 ], image analysis [ 33 , 37 ], image watermarking [ 27 ] etc. Among the
several moment types, the orthogonal moments [ 4 ] constitute the most prominent
moment features (discrimination features based on moments) due to their minimum
redundancy and high reconstruction capabilities. Additionally, their inherent proper-
ties staying invariant under common geometrical transformations (rotation, scaling,
translation, flipping) or incorporating such invariances through coordinates transfor-
mation, give them all the desirable advantages for any invariant pattern recognition
task.
However, a common drawback is the absence of a prior knowledge regarding
the number and the suitability of the used moment features being controlled by
adjusting the order of the orthogonal polynomial used as kernel function. A common
practice is to compute all the moment features up to a certain order and to apply
the entire set of moments as discriminative features. This is an “ad hoc” practice in
some sense, since the significance of each moment component in discriminating the
patterns of the application is not taken into account. A possible solution to this issue
is the application of an additional process that selects, from a large pool, the moment
features that best perform in terms of recognition accuracy.
The aforementioned issue, of the used moments' appropriateness, constitutes the
main subject of this chapter. Initially, the main properties of some representative
moment features and their representation capabilities are discussed in Sect. 13.2 .
Section 13.3 focuses on the justification of the need for selection of the moment
features that better describe the distinctive characteristics of the patterns. The selec-
tion of moment features by applying two different types of selection algorithms, a
Genetic Algorithm and the Relief algorithm, is presented in Sect. 13.4 . An extensive
experimental study with four benchmark pattern recognition datasets and selected
moment features subsets aims to justify the initial assertions in Sect. 13.5 . Finally,
Sect. 13.6 summarizes and discusses the main conclusions.

13.2 Image Moment Features

Geometric moments were the first type of moments introduced in image analysis and
pattern recognition [ 9 ]. These moments constitute the projection of the image on the
monomial x n y m of

+

( n m ) th order. However, the geometric moments suffer from