New Banknote Number Recognition Algorithm Based
on Support Vector Machine
Shan Gai
School of Information Engineering
Nanchang Hangkong University
Nanchang, China
gaishan886@163.com
Guowei Yang
School of Information Engineering
Nanchang Hangkong University
Nanchang, China
ygw_ustb@163.com
Sheng Zhang Minghua Wan
School of Information Engineering School of Information Engineering
Nanchang Hangkong University Nanchang Hangkong University
Nanchang, China Nanchang, China
zhangsheng168@nchu.edu.cn wanminghua@nchu.edu.cn
Abstract—Detecting the banknote serial number is an
important task in business transaction. In this paper, we propose
a new banknote number recognition method. The preprocessing
of each banknote image is used to locate position of the banknote
number image. Each number image is divided into non-
overlapping partitions and the average gray value of each
partition is used as feature vector for recognition. The optimal
kernel function is obtained by the semi-definite programming
(SDP). The experimental results show that the proposed method
outperforms MASK, BP, HMM, Single SVM classifiers.
Keywords—Support Vector Machine; Multiple Kernel Learning;
Semi-definite Programming; Banknote Number Recognition
I. INTRODUCTION
With the development of various banking services and
digital imaging technologies, automatic method for paper
currency recognition and counterfeit recognition are required.
Therefore, finding an effective method to detect counterfeit
banknotes is an demanding task in business transactions.
The serial number of banknote is the key characteristic for
recognition. Each banknote has only one serial number which
can identify the feature of the banknote. The recognition of
banknote is an effective way to solve the problem of
counterfeit banknote recognition. Several methods are
proposed for banknote serial number recognition. Takeda et al
[1] applied mask and neural networks techniques to the
banknote serial number recognition. Frosini et al [2] has used
neural networks to the recognition system. Alex et al [3]
proposed a novel connectionist system which is used to
unconstrained character recognition. Pirlo et al [4] used a new
class of membership function for non-based classification.
Anne et al [5] extracted contextual information from HMM
modeling. Bharath et al [6] proposed two techniques: Iexicon
driven and Iexicon free for recognition. Gon et al [7] allied
projection to locate the position of banknote serial number.
Duan et al [8] proposed a method via location of serial number
region of the banknote.
Vapnik et al [9-13] proposed support vector machine (SVM)
which has been demonstrated to be an effective tool for
classification problem. The main idea of SVM is to construct
hyper-plane which maximized the margin of separation
between classes. It is very important to determine the kernel
functions and hyper parameters of the SVM [14-16] are
proposed by several researchers which can solve the single
kernel SVM. Domenico et al [17] transformed the multiple
kernel SVM into a semi-definite programming (SDP) problem.
Feng Cai et al [18] applied sequential minimal optimization
(SMO) algorithm to learn the optimal multiple kernel weights.
The banknote number recognition system used multiple
kernel SVM is proposed in this paper. The proposed SVM
method is applied to this system. After locating and segmenting
the number of the banknote, the each character is dividing into
non overlapping regions. Each region respond to one kernel,
then the SMO algorithm is used to learn the optimal weight of
multiple SVM. The SDP is applied to reduce the dimension of
the extracted feature vector.
The paper is organized as follows. The review of the SVM
is described in SectionĊ. In Section ċ, the proposed SVM
method is presented. The number recognition system is
proposed in Section Č. The experimental results and analysis
are given in Section č. Finally, the conclusion is presented in
Section Ď.
II.
S
UPPORT VECTOR MACHINES
In this section, we will briefly review the theory of the
SVM. SVM is a kernel method which can solve the standard
learning problem with labeled training data. The maximum
margin hyper-plane is found in the feature space. The decision
function of the SVM is
hx W x b
M
, where
M
is a
mapping from input space to feature space. The
parameters and are calculated by the following equation
respectively.
w b
2013 Second IAPR Asian Conference on Pattern Recognition
978-1-4799-2190-4/13 $26.00 © 2013 IEEE
DOI 10.1109/ACPR.2013.115
176
2013 Second IAPR Asian Conference on Pattern Recognition
978-1-4799-2190-4/13 $26.00 © 2013 IEEE
DOI 10.1109/ACPR.2013.115
176
2013 Second IAPR Asian Conference on Pattern Recognition
978-1-4799-2190-4/13 $26.00 © 2013 IEEE
DOI 10.1109/ACPR.2013.115
176