Professional Documents
Culture Documents
a r t i c l e
i n f o
Article history:
Received 27 February 2011
Available online 11 January 2012
Communicated by D. Coeurjolly
Keywords:
Image thresholding
Iterative selection
Discriminant analysis
Minimum error thresholding
Mixture of Gaussian distributions
Otsus method
a b s t r a c t
There are close relationships between three popular approaches to image thresholding, namely Ridler
and Calvards iterative-selection (IS) method, Kittler and Illingworths minimum-error-thresholding
(MET) method and Otsus method. The relationships can be briey described as: the IS method is an iterative version of Otsus method; Otsus method can be regarded as a special case of the MET method. The
purpose of this correspondence is to provide a comprehensive clarication, some practical implications
and further discussions of these relationships.
2012 Elsevier B.V. All rights reserved.
1. Introduction
In this correspondence, we aim to discuss the close relationships between three approaches to image thresholding, namely
Ridler and Calvard (1978)s or Trussell (1979)s iterative-selection
(IS) method, Kittler and Illingworth (1986)s minimum-errorthresholding (MET) method and Otsu (1979)s method.
With assumptions of bimodal or multimodal probability density
functions of grey levels x, these three approaches are widely used
in practice and highly cited by scientic publications. They are covered in some popular textbooks such as that written by Gonzalez
and Woods (2002, 2008). The MET method is ranked as the best
in a comprehensive survey of image-thresholding methods by
Sezgin and Sankur (2004) recently. Otsus method is implemented
as the default approach to image thresholding in some commercial
and free software such as MATLAB (The MathWorks, Inc.) and
GIMP (www.gimp.org).
The popularity of all three approaches is not a coincidence.
Recently, Xu et al. (2011) prove that, for image binarisation (or
two-level thresholding), Otsus optimal threshold is the threshold t
that equals the average value of the two class means, denoted by
l0(t) and l1(t), for the two classes separated by t. That is,
t = {l0(t) + l1(t)}/2. This result is in fact the iterative rule underlying the IS method. Such a link between the IS method and Otsus
Corresponding author. Tel.: +44 20 7679 1863; fax: +44 20 3108 3105.
E-mail addresses: jinghao.xue@ucl.ac.uk (J.-H. Xue), zhang-yj@tsinghua.edu.cn
(Y.-J. Zhang).
0167-8655/$ - see front matter 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.patrec.2012.01.002
method has also been built by other studies, such as Reddi et al.
(1984) and Magid et al. (1990).
Indeed, as we shall clarify more comprehensively in this correspondence, these three approaches are closely related to each
other: briey speaking, the IS method is an iterative version of
Otsus method; Otsus method can be regarded as a special case
of the MET method.
We shall show that, between the IS method, Otsus method and
the MET method, the links can be readily built from the perspective
of using a Gaussian-mixture distribution to model the grey-level
distribution of an image, as indicated by Kurita et al. (1992) and
Kittler and Illingworth (1986), among others. Such a perspective
is different from, and complementary to, that of Reddi et al.
(1984), Magid et al. (1990) and Xu et al. (2011).
In this context, although this correspondence may mainly revisit some results from various classical literature, our intentions are
twofold. First, we intend to provide the practitioners with a more
comprehensive clarication and some practical implications of
the close relationships between these three popular approaches.
Secondly, we intend to encourage further discussions about effectively applying, extending and evaluating the established imagethresholding approaches.
2. Relationships between the three approaches
Here we only consider image binarisation, but the discussions
presented in the following sections can be readily generalised to
multi-level thresholding.
794
log
py 1jx
p1 pxjy 1
log
0:
py 0jx
p0 pxjy 0
Let us assume that, for each class y, the class-conditional distribution p(xjy) is a Gaussian Nly ; r2y distribution, where ly and r2y
are the mean and variance for class Cy . It follows that Eq. (1)
becomes
p1
r1 x l1 2 x l0 2
log log
0:
p0
r0
2r21
2r20
l0 l1
2
(
t argmaxt max
Xt
"
Xt
"
N
X
#)
logfpyi ; tpxi jyi ; tg
i1
where the parameters Xt p0 t; l0 t; r20 t; l1 t; r21 t are
estimated by their maximum-likelihood estimators (i.e. their sample estimators). This methodology can be traced back to Kittler
and Illingworth (1986).
Under further assumptions that p1(t) = p0(t) and r21 t r20 t,
Kurita et al. (1992) also show that Otsus method is equivalent to
the search for the threshold t that provides the largest maximum
log-likelihood based on p(xjy; t). That is,
t argmaxt max
strategy makes the latter more robust. The strategy and its positive
inuence may be justied by the fact that the IS method can be
viewed as a special case of the iterative MET method.
Furthermore, both methods can be derived from the discriminant
function logfpC1 jx=pC0 jxg 0. From this perspective, the iterative MET method is based on the Gaussian-based quadratic discriminant analysis (QDA), and the IS method is based on a special case of
the Gaussian-based linear discriminant analysis (LDA). The QDA reverts to the LDA, if equal variances are assumed; in Eq. (2), the quadratic term x2 disappears when r20 r21 . Therefore, some practical
guidelines and pitfalls associated with the LDA and the QDA may
be applied to the IS method and the iterative MET method; this merits further empirical investigation, although beyond the scope of this
correspondence.
(
N
X
)#
log pxi jyi ; t
i1
where Xt l0 t; l1 t; r2W t , in which r2W t, denoting r21 t
2
and r0 t, is estimated by the within-class variance. In fact, Eq.
(5) can be obtained through the degeneration of Eq. (4), under those
further assumptions that p1(t) = p0(t) and r21 t r20 t.
In short, Otsus method can be viewed as a special case of the
MET method, in which case equal sizes and equal variances are further assumed for the two classes.
795
6000
Frequency
2000
3000
1000
Frequency
5000
10000
tOtsu
50
100
150
200
tOtsu
50
Grey level
100
150
Grey level
Fig. 1. Otsus binarisation of simulated data for two Gaussian classes. Left-hand panel: two classes with equal variances but distinct sizes (5%:95%); right-hand panel: two
classes with equal sizes but distinct variances r20 256; r21 16. Otsus thresholds tOtsu, indicated by solid lines, split the class with a larger size (see the left-hand panel),
and bias towards the class with a larger variance (see the right-hand panel).
p0 t
pxdx;
l0 t
p0 t
Z
0
p1 t
pxdx:
l1 t
xpxdx;
p1 t
7
Z
3. Further Discussions
xpxdx:
dr2B t=dt 2l0 ttpt l20 tpt 2l1 ttpt l21 tpt:
Given p(t) > 0 and l0(t) l1(t), with simple algebra, we can obtain
t = {l0(t) + l1(t)}/2 from setting dr2B t=dt 0, as shown in Reddi
et al. (1984).
6000
4000
0
2000
Frequency
8000
10000
796
tOtsu
100
150
200
1.0
0.8
0.6
0.4
0.2
0.0
Grey level
80
100
120
140
160
180
200
180
200
Threshold t
160
140
120
80
100
Final threshold
180
200
80
100
120
140
160
Initial threshold
Fig. 2. Binarisation of simulated data for two Gaussian classes. Top panel: Otsus
threshold, tOtsu, for two classes with equal variances but highly-distinct sizes
(2%:98%). Middle panel: the within-class variance r2W t used by Otsus method for
searching for an optimal threshold tOtsu; rescaled for illustrative purposes. Bottom
panel: nal thresholds obtained by the IS method versus corresponding initial
values of the thresholds.
4. Summary
In this correspondence, we have provided a comprehensive
clarication of the close relationships between three popular image-thresholding approaches. That is, in short, Ridler and Calvards
IS method is an iterative version of Otsus method; Otsus method
can be regarded as a special case of Kittler and Illingworths MET
method. It was our expectation that such a clarication could help
the practitioners to understand more comprehensively the characteristics, thresholding performances and pitfalls of these approaches, and thus facilitate the application, extension and
evaluation of them.
Acknowledgements
The authors are grateful for the referees and the Area Editors
constructive comments, in particular those on the unimodal thresholding and the evaluation of image-thresholding methods.
References
Bazi, Y., Bruzzone, L., Melgani, F., 2007. Image thresholding based on the EM
algorithm and the generalised Gaussian distribution. Pattern Recognition 40 (2),
619634.
Chapelle, O., Schlkopf, B., Zien, A. (Eds.), 2006. Semi-Supervised Learning. The MIT
Press, Cambridge, MA.
Dong, L., Yu, G., Ogunbona, P., Li, W., 2008. An efcient iterative algorithm for image
thresholding. Pattern Recognition Letters 29 (9), 13111316.
Fan, S.-K.S., Lin, Y., Wu, C.-C., 2008. Image thresholding using a novel estimation
method in generalised Gaussian distribution mixture modelling.
Neurocomputing 72 (1-3), 500512.
Fernndez-Garca, N.L., Medina-Carnicer, R., Carmona-Poyato, A., Madrid-Cuevas,
F.J., Prieto-Villegas, M., 2004. Characterization of empirical discrepancy
evaluation measures. Pattern Recognition Letters 25 (1), 3547.
Gonzalez, R.C., Woods, R.E., 2002. Digital Image Processing, second ed. Prentice
Hall, Upper Saddle River, NJ.
Gonzalez, R.C., Woods, R.E., 2008. Digital Image Processing, third ed. Pearson
Prentice Hall, Upper Saddle River, NJ.
Jiulun, F., Winxin, X., 1997. Minimum error thresholding: A note. Pattern
Recognition Letters 18 (8), 705709.
Kittler, J., Illingworth, J., 1985. On threshold selection using clustering criteria. IEEE
Transactions on Systems, Man, and Cybernetics SMC-15 (5), 652655.
Kittler, J., Illingworth, J., 1986. Minimum error thresholding. Pattern Recognition 19
(1), 4147.
Kittler, J., Illingworth, J., Fglein, J., 1985. Threshold selection based on a simple image
statistic. Computer Vision, Graphics, and Image Processing 30 (2), 125147.
Kurita, T., Otsu, N., Abdelmalek, N., 1992. Maximum likelihood thresholding based
on population mixture models. Pattern Recognition 25 (10), 12311240.
Lee, H., Park, R.-H., 1990. Comments on An optimal multiple threshold scheme for
image segmentation. IEEE Transactions on Systems, Man, and Cybernetics 20
(3), 741742.
Levine, M.D., Nazif, A.M., 1985. Dynamic measurement of computer generated
image segmentations. IEEE Transactions on Pattern Analysis and Machine
Intelligence 7 (2), 155164.
Magid, A., Rotman, S.R., Weiss, A.M., 1990. Comment on Picture thresholding using
an iterative selection method. IEEE Transactions on Systems, Man, and
Cybernetics 20 (5), 12381239.
Medina-Carnicer, R., Madrid-Cuevas, F.J., 2008. Unimodal thresholding for edge
detection. Pattern Recognition 41 (7), 23372346.
Medina-Carnicer, R., Muoz-Salinas, R., Carmona-Poyato, A., Madrid-Cuevas, F.J.,
2011. A novel histogram transformation to improve the performance of
797