. INTRODUCTION
HE business information system has an important rule in
organizations. It is a support tool for business activities
to reach organization goals. One of the main components of
the business system is business transaction data, which can
be collected from several sources. In a practical way, most of
sources of the transaction data usually are gotten from
printed documents. Hence, a procedure of transforming
printed documents to a computer-understandable form is a
need. The OCR software is an automatic tool for that
procedure. Nowadays, Thai OCR software is not widely used
in business applications. Since its obtained results is slightly
low.
In decade past, many researchers have been introduced
several Thai OCR techniques that cover in area of local
feature extraction techniques, classifier techniques, and
automatic word correction techniques such as [1],[2],[3],[4]
and [5]. However, the existing techniques are not successful.
According to a set of Thai alphabet shown in Figure 1,
several Thai characters are too similar such as “”, “
”,
and “”. In particularly, as the images of similar characters
are obtained from a poor paper, the image characters may be
more confused as shown in Figure 2. So, the character
recognition procedure with the local feature extraction
techniques may not handle the case of low quality character
images. Furthermore, the recognition procedure with word
correction techniques also may not handle when a number of
incorrect characters obtained from the recognition procedure
are about 3-5 characters/word. Hence, a skilful classifier is a
need for the recognition procedure. Nowadays, Thai OCR
researchers have introduced several classifiers to the
recognition procedure. However, they attentively apply an
individual classifier to the recognition procedure. In this
paper, we will concentrate on using the combination of
multiple classifiers to achieve better the accuracy of
classification rate. Furthermore, to reduce matching times of
the classifiers, a set of prototypes obtained by the SGNG
algorithm is employed to roughly classify an unknown
pattern, when the position of the pattern is located around
high density regions of a training dataset.
The paper is organized as follows: section II discusses the
details of the proposed recognition system. The experimental
results with respect to a feature extraction parameter and the
performance of the proposed system are presented in section
III. Finally, the conclusion is given in the last section.