10-04-2017, 07:58 PM
Abstract
This paper demonstrates some consideration results
on feature extraction and selection for handwritten
Japanese Hiragana characters using independent
component analysis (ICA). In some ICA
algorithms where whitening of input signals is introduced
as preprocessing, one can consider that
the process of feature extraction using ICA consists
of two types of transformations: one is the transformation
from an input image to its principal components
(PCs), and the other is the transformation
from PCs to independent components (ICs). From
this fact, two types of feature selection can be applied
to outputs of these transformations (i.e. PCs
and ICs). Furthermore, as criteria of useful features,
cumulative proportion can be adopted in the
former type of feature selection, and kurtosis can
be adopted in the latter. Thus, we present ve different
feature selection methods in this paper. To
discuss the e ectiveness of these methods, recognition
experiments using hand-written Japanese Hiragana
characters are carried out. As a result, we
show that a hybrid method, in which feature selection
is carried out for ICs as well as for PCs,
has attractive characteristics if small dimensions of
feature vectors are preferred in classi cation.
1 Introduction
Recently, independent component analysis (ICA)
has been widely noticed as a decorrelation
technique based on higher-order moment of input
signals[1]. Using such characteristics, ICA has been
so far applied to problems of blind signal separation
such as sound/image separation and EEG signal
separation. On the other hand, feature extraction
of images and sounds has been also focused as one of
prominent applications of ICA[2, 3, 4, 5]. Bartlett
& Sejnowski extracted feature vectors from images
of human faces with ICA, and showed that these
feature vectors had greater viewpoint invariance
for human faces as compared with PCA (principal
component analysis) ones[6]1. Since PCA decorrelates
only the second order statistics of input signals,
this result indicates that higher-order features
are useful for capturing invariant features of face
patterns as well as conventional second-order features.
Such invariant characteristics of ICA feature
vectors might be attractive for other pattern recognition
problems.
In our previous work, ICA feature vectors are
utilized for recognizing hand-written characters in
order to study the usefulness of ICA as a feature extraction
method[7]. An ICA feature vector is given
as a coe cient vector of the bases that is obtained
by applying an ICA algorithm to a training set of
character images. From the experimental results,
we showed that the recognition accuracy greatly
depended on input dimensions; in general, small
dimensions of inputs tend to attain good classi cation
performance. In this work, input dimensions
were changed with the size of subimages, which
are separated by imposing small size of windows
to a character image. Although such reduction of
input dimensions is actually e ective, one can say
that only local features of a character image are
considered in this approach. To achieve higher performance,
we should use global features as well as
local ones; that is, we should reduce dimensions by
selecting useful features extracted from the whole
of a character image.
In this paper, we will study feature selection
methods in order to obtain high-performance ICA
feature vectors through some recognition experiments
for hand-written Japanese Hiragana characters.
In Section 2, we will brie
y refer to an ICA
algorithm and its objective function. In Section 3,
a method of feature extraction using ICA will be
described, then we will propose some criteria for
feature selection. In Section 4, recognition experiments
will be carried out, and we will discuss what
criteria are suitable for feature selection.
Download full report
http://googleurl?sa=t&source=web&cd=1&ve...be-u.ac.jp%2F ozawasei%2Fpub%2Ficonip00.pdf&ei=SgUxTr_oL8OOsAKW4qDpCg&usg=AFQjCNEpeaEsVILLh_-ENKQvxIGz0WRp5g&sig2=TfWvc_-V4D4IOHByUN6vKQ