See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220632416

A Survey of Thresholding Techniques

Article  in  Computer Vision Graphics and Image Processing · February 1988

DOI: 10.1016/0734-189X(88)90022-9 · Source: DBLP

CITATIONS READS

2,223 5,614

3 authors:

Prasanna Sahoo Sasan Soltani

University of Louisville 3 PUBLICATIONS   2,243 CITATIONS   
174 PUBLICATIONS   7,741 CITATIONS   
SEE PROFILE
SEE PROFILE

Andrew Wong
University of Waterloo
229 PUBLICATIONS   7,718 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Segmentation of Color Images using Information Theory and PSO View project

Pattern discovery for large mixed-mode database View project

All content following this page was uploaded by Sasan Soltani on 06 November 2018.

The user has requested enhancement of the downloaded file.

COMPUTER VISION, GRAPHICS, AND IMAGE PROCESSING 41, 233-260 (1988)

SURVEY

A Survey of Thresholding Techniques*

P. K. SAHOO, S. SOLTANI, AND A. K. C. W O N G
Department of SystemsDesign Engineering, University of Waterloo, Waterloo, Canada N2L 3GI

AND
Y. c. CHBN
Department of Electrical Engineering, University of Waterloo, Waterloo, Canada N2L 3GI
Received July 7,1984; accepted July 9,1987

In digital image processing, thresholding is a well-known technique for image segmentation.

Because of its wide applicability to other areas of the digital image processing, quite a number
of thresholding methods have been proposed over the years. In this paper, we present a survey
of thresholding techniques and update the earlier survey work by Weszka (Comput. Vision
Graphics 62 Image Process 7, 1978, 259-265) and Fu and Mu (Pattern Recognit. 13, 1981,
3-16). W e attempt to evaluate the performance of some automatic global thresholding
methods using the criterion functions such as uniformity and shape measures.The evaluation
is based on some real world images. 0 1988 Academic press, IX.

1. INTRODUCTION
A popular tool used in image segmentationis thresholding. This paper represents
a survey of a variety of thresholding (also known as binarization) techniques
including both global and local thresholding. For the sake of discussion, global
techniques are further classified as: point-dependent and region-dependent tech-
niques. Several global thresholding methods are examined in detail to evaluate their
performance based on a given set of test images. In this paper, we endeavor to
update the earlier survey papers by Weszka [44] and Fu and M u [13]. W e also
attempt to evaluate the performance of several global thresholding methods using a
shape measure and a region uniformity measure as criterion functions. The evalua-
tion is conducted on a set of images digitized from photographs of a portrait and
natural scenes.Although our primary goal is to present a survey of bilevel threshold-
ing methods, extensions of these methods to m u ltithresholding problems are also
treated in this paper.
Let N be the set of natural numbers, (x, y) be the spatial coordinate of a
digitized image, and G = (0, 1, . . . , I - 1) be a set of positive integers representing
gray levels. Then, an image function can be defined as the m a p p ing f : N x N + G .
The brightness (i.e., gray level) of a pixel with coordinate (x, y) is denoted as
f(K Y)-
Let t E G be a threshold and B = {b,, b,} be a pair of binary gray levels and
b,,b, E G . The result of thresholding an image function f(. , a) at gray level t is a

*This work was supported by the Natural Sciencesand Engineering Reearch Council, Canada, under
Grants A4716 and G1367.
233
0734-189X/88 $3.00
Copyright 0 1988 by Academic Press, Inc.
All rights of reproduction in any form reserved.
234 SAHOO ET AL.

binary image function f, : N x N + B, such that

In general, a thresholding method is one that determines the value t* of t based

on a certain criterion. If t* is determined solely from the gray level of each pixel,
then the thresholding method is point-dependent. If t* is determined from the local
property (e.g., the local gray level distribution) in the neighborhood of each pixel,
then the thresholding method is region-dependent. A global thresholding technique
is one that thresholds the entire image with a single threshold value, whereas a local
thresholding technique is one that partitions a given image into subimages and
determines a threshold for each of these subimages.
Let the number of pixels with gray level i be ni. Then the total number of pixels
in a given image is
I-1
n = C ni.
r=O
The probability of occurrence of gray level i is defined as

p,= “I.
n
Also, by convention, the gray level 0 is the darkest and the gray level 1 - 1 is the
lightest.
2. GLOBAL THRESHOLDING: POINT DEPENDENT TECHNIQUES

A. p-tile Method
One of the earliest thresholding methods is the p-tile method [lo]. In this method,
an image is assumed to consist of dark objects in a light background. By assuming
that the percentage of the object area is known, the threshold is defined as the
highest gray level which maps at least (100 - p)% of the pixels into the objects in
the thresholded image. For example, suppose an object occupies 20% of an image,
then the image should be thresholded at the highest gray level that allows at least
20% of the pixels to be mapped into the object. Obviously, this method is not
applicable to images whose object area is not known.
B. Mode Method
For images with distinct objects and background, the histogram of the gray levels
will be bimodal. In this case, a threshold can be chosen as the gray level that
corresponds to the valley of the histogram. The technique is called the mode method
[31]. Though this method is simple, it cannot be applied to images with extremely
unequal peaks or to those with broad and flat valleys.
C. Ostu Method
This method, as proposed in [27], is based on discriminant analysis. In this
method, the threshold operation is regarded as the partitioning of the pixels of an
 SURVEY OF THRESHOLDING TECHNIQUES 235

image into two classes C,, and C, (e.g., objects and background) at gray level t. That
is, C, = {O,l,. . . , t}andC,={t+l,t+2,...,1-l}.Leta~,u&andc$bethe
within-class variance, between-class variance, and the total variance, respectively.
An optimal threshold can be determined by m inimiz,ing one of the following
(equivalent) criterion functions with respect to t:

4
q=<, and K=y.
UT UW

Of the above three criterion functions, n is the simplest. Thus, the optimal
threshold f* is

t* = ArgMinn,
tEG

where
I-1 I-1
U$= C(i-jJT)*Pi, PT= zipi?
i=O i=O

t
wO= CPi, 01 = 1 - wg,
i=O

PT- CL,
po=II, p‘= i ip,.
I-%= l-oo’
*0 i=O

D. Histogram Concavity Analysis Method

For images with distinct objects and background it is possible to select the
threshold from the gray level histogram using the mode method. For some images
where valleys may not be found in their gray level histogram, it is often possible to
define a good threshold at the “shoulder” of the histogram. Since both valleys and
shoulders correspond to the concavities in the histogram, a threshold can be
determined by analyzing the concavity of the histogram [35].
Let HS be a histogram defined over a set of gray levels go, g,, . . . , g,-,. Denote
the height of the histogram at these gray levels as h( go), h( gl), . . . , h(g,-,), where
h( gi) # 0 for all i. Thus, HS can be regarded as a two-dimensional region.
To determine the concavities of HS, its convex hull is constructed. It is the
smallest convex polygon m which contains HS. The concavities of HS are
determined from the set-theoretic difference HS - B. Let h ( gi) be the height of iE
at gray level gi. Possible threshold values are gray levels at which & ( gi) - h( gi) has
local maxima. However, not all these maxima are good candidates for thresholding
because large concavities can also be introduced by noise spikes. Rosenfeld and
De La Torre [35] call such concavities spurious. In order to eliminate maxima caused
by spurious concavities, a balance measure
236 SAHOO ET AL.

is introduced. Ei measures the balance of the histogram about the gray level g,. For
spurious concavities, which usually lie only on one side of the histogram, the values
of E; will be small. Thus, spurious concavities can be eliminated by ignoring
maxima of h - h when Ei is small. The remaining maxima indicate possible
locations for thresholding, but they may not be optimal. Other gray levels near these
maxima should also be considered for possible improvement.

E. Entropic Methods
In these recently developed methods, the gray level histogram is considered as a
I-symbol source. The optimal threshold is obtained by applying information theory.

(i) Pun Methods

In this subsection, two recently developed algorithms proposed by Pun [32,33]
will be discussed.
Let t be the value of the threshold and define two a posteriori entropies [l]

where H,: and H; can be regarded, respectively, as measures of the a posteriori

information associated with the black and white pixels after the thresholding.
Knowing the a priori entropy of the gray level histogram, Pun [32] proposes an
algorithm to determine the optimal threshold by maximizing the upper bound of the
a posteriori entropy

H’ = H,: -I- H;.

Pun [32] has shown the maxir&in g H’ is equivalent to maximizing the evaluation
function

+ l--
4
[ 1 log,U
4
- PA
log,m~~p,+l~.--~ p,-d

with respect to t, where

Ht = - i p;log,p;,
i=o
I-l
HT = - C PihePi>
i=O

‘t’ CPi’
i=O
 SURVEY OF THRESHOLDING TECHNIQUES 237

In his second algorithm, Pun [33] proposes the use of an anisotropy coefficient, (Y,
in thresholding, where

i=o
a= I-1
1;oPi10gt3 Pi ’

and m is the smallest integer such that

5 p; 2 0.5.
i=O

The optimal threshold t* is chosen such that

However, Kapur et al. [16] have shown that this algorithm will always threshold an
image with t* 2 m, thus introducing unnecessary bias.

(ii) Kapur, Sahoo, and Wong Method

In this method [16], two probability distributions (e.g., object distribution and
background distribution) are derived from the original gray level distribution of the
image as follows:

--PO Pl Pt
P, ’ P, ’. * . ’ p,

and

--Pt+l Pt+2 P/-l

1 - P, ’ 1 - P, ‘. ..’ 1 - P, ’

where t is the value of the threshold and P( = &opi. Define

H,(t) = - i Flog, ; )
i=o t i ti
l-l
H,(t) = - c Alog, & .
i=t+l ’ - ‘t i ti

Then the optimal threshold t* is defined as the gray level which maximizes
H,,(t) + H,,,(t), that is,
238 SAHOO ET AL.

(iii) Johannsen and Bille Method

This method [15] uses the entropy of the gray level histogram of the digital image.
Essentially, it divides the set of gray levels into two parts so as to minimize the
interdependence (in information theoretic sense) between them. The Joharmsen and
Bille method chooses the threshold value t* from the relation

t* = Argzg{S(t) + g(t)},

where

and

i II
l-l

? i=:lpi .

i=l

F. Moment-Preserving Method
In this method [40], the threshold values are computed deterministically in such a
way that the moments of an image to be thresholded are preserved in the output
(binary) image. This i” moment mi is calculated as

m, = i ‘$ g’h(g), i = 1,2,3,
g-0

where n is the total number of pixels in the image. The threshold value t* is
obtained from the gray level histogram of the image by choosing t* as the PO-tile,
where p. is given by
z - ml
PO = (c; - &,y2

and

m1m3 - rn$ ml1232 - m3

co = 2 Cl = 3
m2 - rnf m2 - rnf
z+
0 c; - 4coy2 - Cl).
 SURVEY OF THRESHOLDING TECHNIQUES 239

G. Minimum Error Method

In the minimum error method [19], the gray level histogram is viewed as an
estimate of the probability density function P(g) of the mixture population
comprising of the gray levels of the object and background pixels. It is usually
assumed that each of the two components p(gli) of the mixture is normally
distributed with mean pi and standard deviation ai and a priori probability Pi, that
is

P(g) = i P,P(gli)9
r=l

where

PW) = & exp( - (gi-$i)).

I
The threshold value can be selected by solving the quadratic equation

(g -0: PA2+ 1ogJJ: - 2log,P, =

(g -
4
P212
+ 1ogJJ; - 2log,P,.

However, the parameters pi, u/, and Pi (i = 1,2) of the mixture density p(g)
associated with an image to be thresholded are not usually known. To overcome the
difficulty of estimating these unknown parameters, Kittler and Illingworth [19]
introduced a criterion function J(t), which is given by

J(t) = 1 + 2{plw%e%w + p2(tbgea,(t)~

-2{wbg,pl(t) + p20)1%920)~~

where
t I-1

40) = c h(g), P,(t) = c h(g)

g=o g=t+l

l%(t)=

and
l-l

c (g - dt))2h(d
&) = i g=t+l
P,(t)
240 SAHOO ET AL.

The optimal threshold is obtained by minimizing J(t), that is, by finding

t* = Arg%zJ(t).

3. GLOBAL THRESHOLDING: REGION DEPENDENT TECHNIQUES

A. Histogram Transformation Methods

The methods described here do not select a threshold directly. Rather, they
transform the gray level histogram of the image into one with deeper valleys and
sharper peaks so that the mode method described in Section 2B can be applied to
determine the threshold. A common feature among these methods is that a new
histogram is obtained by weighting the pixels of the image according to the local
property of the pixels. Also, these methods assume that each of the images
considered consists of objects and background, both of which have a unimodal gray
level distribution.
Mason et al. [23] propose the use of the edge operator (e.g., the Laplacian, the
Roberts cross, etc.) for weighting. According to their method, the values of the edge
operator are small for pixels in homogeneous regions and these pixels are given
more weight. However, the values of the edge operator are large for pixels in the
neighborhood of an edge and these pixels are given less weight. The weight can be
computed as l/(1 + A’), where A is the edge value at a given pixel. As a result of
this weighting process, the new gray level histogram will have sharper peaks and
deeper valleys.
In 1965, Katz [17] points out that since the pixels in the neighborhood of an edge
have higher edge values, the gray level histogram for these pixels should have a
single peak at a gray level between the object and the background gray levels. This
gray level is, therefore, a suitable choice of the threshold value. This has also been
studied by We&a and Rosenfeld [46]. Several variations of the above method have
been proposed [45,43,49]. Weszka and Rosenfeld [48] unifies them by using a gray
level versus edge value scatter plot.
Another method for histogram transformation that draws our attention is the
quadtree method [51]. This method is based on the fact that the standard deviation
of the gray levels in a homogeneous region is small, whereas that in a nonhomoge-
neous region is high. Thus, regions whose gray level standard deviations are high
can be divided into smaller, homogeneous regions. Starting from the original image,
the quadtree method divides the image into quadrants if its gray level standard
deviation exceeds a predefined value. The process is then repeated for each quadrant.
This yields a decomposition of the original image into blocks with lower gray level
standard deviation so that it is reasonable to approximate the gray level of each
block by its mean gray level. The resulting image is called the Q-image. Because of
the homogeneity of each block, the gray level histogram of the Q-image will have
sharper peaks and deeper valleys.
B. Methods Based on Second-Order Gray Level Statistics
One of the drawbacks of the point-dependent thresholding methods is that they
depend solely on the first-order gray level statistics (i.e., the histogram) of the image.
The methods to be described here are based on the second-order gray level statistics
of the image.
 SURVEY OF THRESHOLDING TECHNIQUES 241

(i) Co-occurrence Matrix Method

The co-occurrence matrix is introduced by Haralick for texture analysis [14]. In
general, a co-occurrence matrix MCfl,+) is one whose entries are the relative
frequency of occurrence for two neighboring pixels with gray levels i and j,
separated by distance d and with orientation +. Obviously, the number of such
matrices can be quite large, depending on the choice of d and 9. In Ahuja and
Rosenfeld [2], the gray level co-occurrence matrix is defined as

that is, the (i, j) element of M is the frequency that gray level i occurs as a
4-neighbor of gray level j.
Because of homogeneity, pixels interior to the objects or background should
contribute mainly to the near-diagonal entries of M . Also, pixels near an edge
should contribute mainly to the off-diagonal entries of M because of the gray level
change near an edge. Therefore, the matrix M can be used to define two new
histograms:
(a) A histogram based on the near-diagonal entries of M . This histogram
should have a deep valley between the object and the background gray levels.
(b) A histogram based on the off-diagonal entries of M . This histogram should
have a sharp peak between the object and the background gray levels.
A threshold for the image can then be chosen from the gray level range in which
the valley in (a) overlaps with the peak in (b).

(ii) (Gray Level, Local Average Gray Level) Scatter Plot Method
This method [18] is quite similar to the co-occurrence matrix method discussed in
the previous section. In a (gray level, local average gray level) scatter plot, the origin
is defined as the upper left-hand comer, the gray level increases from left to right,
and the local average gray level, taken over a window, increases from top to bottom.
The intensity of a point on this plot is proportional to the occurrence frequency of
the corresponding pair (gray level, local average gray level). In Kirby and Rosenfeld
[18], the local averages are taken over a 3 x 3 square window.
The near-diagonal entries in the scatter plot represent pixels whose local average
gray level is close to the gray level; such pixels are likely to be interior to the objects
or background because of homogeneity. Thus, the histogram of the gray levels of
these pixels should have a deep valley. On the other hand, the off-diagonal entries in
the scatter plot represent pixels near an edge, so the gray level histogram obtained
from these pixels should have a sharp peak. As in the co-occurrence matrix method,
a threshold can be determined from the gray level range in which the valley and the
peak of the above histogram overlap with each other.

C. Deravi and Pal Method

In this method [9], transition matrices similar to the co-occurrence matrix
discussed in Section 3B(i) are used to define two “interaction measures” for
threshold selection. The optimal threshold is determined by m inimizing these
interaction measures.
242 SAHOO ET AL

Using the notations of Section 3B(i), the transition matrices defined in [9] can be
written as

Th = M&O),
T, = JqlJn,2)7
Tub= T,+ T,,.

Any one of the above transition matrices can be used in threshold selection and we
shall refer to the transition matrix as T in the following discussion. It was reported
in [9] that the results obtained with different transition matrices do not differ
significantly.
Let qj be the (i, j) entry of T and t be a threshold dividing the set of gray levels
into two classes: C, = (0, 1, . . . , t}andC,={t+l,t+2,...,1-l}.ThenTcan
be partitioned into four regions defined by the following parameters:
I-1 l-l

a= i &j, b= C C Tj>
i=O j=CJ i=t+l j=t+l
I-1 r
c= i ‘g 71j, d= c c qj.
i=O j=t+l i=f+l j=O

Thus a, b, c, and d represent the total number of transitions within Co, within C,,
from C, to C,, and from C, to C,, respectively. The joint and conditional
probabilities of the transition between Co and C, can be estimated by
c+d
P,(t) =
a+b+c+d’

and

Deravi and Pal [9] call P,(t) and P,(t) interaction measures. They also point out
that P,(t) is similar to the “business measure” in [47] and that P,(t) does not
depend directly on the histogram. The optimal threshold, t*, is obtained by
maximizing either one of the above interaction measures.
D. Relaxation Metho&
The idea of relaxation was introduced by Southwell [38,39] to improve the
convergence of recursive solution for system of linear equations. In image segmenta-
tion, relaxation is applied as follows. The pixels of an image are first probabihsti-
tally classified into “light” and “dark” classes based on their gray levels. Then the
probability of each pixel is adjusted according to the probabilities of the neighbor-
ing pixels. This adjustment process is iterated so that “light probabilities” (resp.
“dark probabilities”) become very high for pixels belonging to the light (resp. dark)
regions. An attractive feature of these relaxation methods is that they are p
processing techniques, as opposed to the sequential techniques discussed so far.
 SURVEY OF THRESHOLDING TECHNIQUES 243

(i) Initial Clksdjchvs;dlkjhv;sldjkhvlassijkation

For the initial classification of the pixels, Rosenfeld and Smith [36] have sug-
gested the following method. Let d and 1 be the darkest and lightest gray levels, and
gj be the gray level of a pixel xi. Then, for xi, let

l - gi
Piqdark= -I-d’

and

gi - d
&i&t- -- 1-d’

Though the above method is simple, it does not work in cases where the object
and background gray levels do not lie on different halves of the gray level histogram.
To avoid this problem, Rosenfeld and Smith [36] suggest another initialization
scheme. Let m be the mean gray level. Then, if gj > m , let

1 1 gi-m
Piqlight=2+-- 2 l-m’

and if gi I m , let

1 lm-g,
p!&&=Z+---
2 m-d’

Fekete et al. [ll] suggest an approach in which they assume the histogram can be
divided into two Gaussian subpopulations so that the gray level distribution can be
written as the sum of two Gaussian distributions. The parameters of these Gaussian
distributions are determined by a method suggested in [7]. They find that faster
convergence of the relaxation process can be obtained with this method.

(ii) Iterative Updating of Probabilities

As previously mentioned, the updating process consists of adjusting the probabih-
ties of each pixel, based on neighboring probabilities. Let A be the set of class labels
(e.g., the classes of dark and light pixels). Then, to direct the class updating, a
compatibility coefficient, rij(X, X’), between a pixel xi with label X E A and
another pixel xi with label X’E A is defined such that

-1 if A and A’are incompatible.

rij(X, A’) = 0 if xi and xj are independent.
1 if X and A’are compatible.
244 SAHOO ET AL

Zucker et al. [52] propose the following equation for updating the probabilities:

p”+‘(h)= P,k(Nl+ dY~)l

&Pm[l + %v)l ’
4fCA)= $ C C rij(A,A’)pj(X’),
x,eN, A’sA

where N, is the 8-neighbor of xi.

However, Pavlidis [29] has shown that p:(X) always changes during the updating,
therefore the above scheme violates the natural expectation that the labeling should
not change if neighboring pixels are independent. Peleg [30] resolves this theoretical
flaw by suggesting the following updating equation:

E. Gradient Relaxation Methods

In gradient relaxation, the optimal labeling scheme is determined by maximizing
a criterion function with gradient optimization.
Let A, and A, be the labels of the classes of light and dark pixels, let
UP,(~), P~@~>I~~i = 61 , . . . ,l - l} be the set of probability vectors associated
with the gray levels and let {[qi(X,), qi(X2)lT, i = O,l,. . . , I - l} be the set of
compatibility vectors, where qi(Xk) = iCX, E N,p&X,), N, is the 8-neighbor of the
pixel xi. Bhanu and Feugeras [4] suggest that the optimal labeling scheme for
images with unimodal gray level distribution can be determined by
I-1
maximizing c(P> ‘= C PTCA)qiCA) subject to
i=o

P;(X) E K:= P(A) = (p(h), P(&>)~P($> 2 0, i P@,) = 1

i j=l

They point out that maximizing C(p) is equivalent to reducing the inconsistency
and ambiguity of the pixel labeling.
Here we suggest a similar method that uses a different criterion function derived
from information theory. In this method the optimal labeling scheme is determined
by
I-1

maximizing *(P) ‘= C z(Pi, 4i)

i=O
subject to pi E K.
 SURVEY OF THRESHOLDING TECHNIQUES 245

The quantity I( pi, qi) is defined as

Pi(‘j)
I( Pi3 4i) = i Pi(‘j ___
j=l e 4iCxj> .
i I

F. Other Methods
Thresholding techniques for optical character recognition systems have also
received much attention. Because of the wide range of print quality distortions over
a single document, a combination of threshold operators is often used, with each
operator sensitive to a different type of distortion. Bartz [3], for example, combines
four linear threshold operators to form a single threshold. An example of these
operators is T = ku + c, where u is the average contrast over previously scanned
characters, and k and c are optimizing parameters. Wolfe [50] suggests a two-step
procedure to deal with the problem of shading in printed characters. During the first
step, the mean gray level of each pixel is computed by averaging the gray levels over
a 4 x 4 window. A pixel X is considered as belonging to a character if its mean
gray level is darker than the mean gray levels of the two pixels whose orientations
differ by 180’ and are distance 8 apart. The second step is similar to the first step,
but a larger window is used.
In Ullmann [41], the threshold of a pixel X is selected based on the gray levels in
a 5 x 5 window, W, of X. Only the pixels which are labeled n contribute to the
threshold decision (Fig. 1).
Two experimental rules, to be described later, are used to determine the threshold
of X. Let n, be the highest (brightest) gray level in W. Then rule (1) is applied if
n x < 40 and rule (2) is applied when n x > 40. The two rules are:
(1) Label X as an object point if for some point n in W, we have g(X) - g(n)
< 7, where g( -) means the gray level of (e) and r is some predefined threshold;
otherwise label it as a background point.
(2) Label X as an object point if for at least one n in W, we have g(X) <
g(n)/p, where 1-1is some predefined constant; otherwise label it as a background
point.
Morrin [25] uses gray level versus gradient value plots to convert gray scale
images with superior resolution and contrast for thresholding. Panda [28] suggests a
method in which the threshold applied at a point depends on both the gray level and
the edge value of that point. Some results and comparisons are given in [28].

0 n n n 0
n n 0 II n
n 0 x 0 n
n n Cl 0 n
0 n n n 0

FIG. 1. 5 x 5 window used in Ullmann [41].

246 SAHOO ET AL.

4. LOCAL THRESHOLDING

In local thresholding, the original image is partitioned into smaller subimages and
a threshold is determined for each of the subimages. This yields a thresholded image
with gray level discontinuities at the boundaries of two different subimages. The
threshold of a region can be determined by either the point-dependent method or
the region-dependent method. A smoothing technique is then applied to eliminate
the discontinuities.
Chow and Kaneko [7, S] suggest the use of a 7 X 7 window for local thresholding.
In their method, the original image is divided into 7 x 7 subimages and a threshold
is computed for each subimage. However, a threshold is not computed for sub-
images with unimodal gray level histogram. Thresholds for such subimages are
interpolated from neighboring subimages. For a bimodal subimage, the threshold is
computed as follows. First the gray level histogram for a subimage is approximated
by a sum of two Gaussian distributions, then the threshold is obtained by minimiz-
ing the classification error with respect to the threshold value. Some experiments on
this method were done in [26].
For X-ray angiograms Fernando and Monro [12] suggest a local thresholding
method. Global thresholding techniques have been found to be unsatisfactory for
these images which are usually unimodal with a very narrow peak. According to this
new method, the original image is partitioned with 16 nonoverlapping subimages
and the entropic thresholding technique of Pun [32] is applied to determine the
threshold value for each of these subimages. Finally, the entire thresholded image is
processed by a low-pass filter to eliminate the gray level discontinuities at the
boundaries of subimages.

5. MULTITHRESHOLDING METHODS

Many global thresholding methods, such as Ostu [27], Pun [32,33], Kapur et al.
[16], moment preserving [40], minimum error [19] can be extended to the case of
multithresholding. In this section we discuss three multithresholding methods which
are not included in the earlier sections of this paper.

A. Amplitude Segmentation Method

This method was proposed by Boukharouba et al. [5] and uses the intrinsic
properties of cumulative distribution function of an image to be thresholded. In this
method, the curvature of the distribution function is examined to obtain informa-
tion regarding the threshold values. The distribution function F(k) at point k is
given by
k

F(k) = “1”
‘h(g) ’
g=o

The curvature of F is then defined by

C(x) = F’(x)[l + (P(x))‘] -3’2,

 SURVEY OF THRESHOLDING TECHNIQUES 247

where F’ and F” are the first and second derivatives of F, respectively. It is pointed
out in [5] that C(x) is noisy and oscillatory and should be smoothed and approxi-
mated for use in thresholding. The zeros of the curvature determines the thresholds
as well as the gray level to be assigned to each class.

B. Wang and Haralick Method

This is a recursive technique [42] for multiple threshold selection on digital
images. In this method, pixels are first classified as edge pixels or nonedge pixels.
Edge pixels are then classified, on the basis of their neighborhoods, as being
relatively dark or relatively light. A histogram of the gray levels is obtained for those
pixels which are edge pixels and relatively dark, and another histogram is obtained
for those pixels which are edge pixels and relatively light. A threshold is selected
based on the gray level intensity value corresponding to one of the highest peaks
from the two histograms. To get multiple thresholds, the procedure is recursively
applied using only those pixels whose intensities are smaller than threshold first and
then using only those pixels whose intensities are larger than the threshold.

C. Uniform Contrast Method

This is a recursive threshold selection method proposed by Kohler [21]. The
method is based on the following idea. The optimum threshold for segmentation of
the image is that threshold which detects more high contrast edges and fewer low
contrast edges than any other threshold [21]. In this method, a histogram of the
average contrast p(t) for each possible threshold t is created and the highest peak
in the histogram corresponds to the optimal threshold. The average contrast p(t) is
calculated from the relation

c(t)
CL(t)= -N(t)
with p(t) = 0 if N(t) = 0 and C(t) is the total contrast detected by threshold t, and
N(t) is the number of edges detected by t. For multithresholding any initial
threshold is first selected and then a new histogram of p(t) is computed by
removing the contribution of the already detected edges by the initial threshold.
This procedure is continued until the maximum average contrast for any threshold
fell below some m inimum average contrast criterion 8 > 1.
6. EXPERIMENTAL RESULTS
Not all the methods surveyed in this paper are automatic in nature (i.e., they
require no human interactions). Some require user’s feedback since unique optimal
threshold is not possible. In local thresholding methods, the image is divided into
smaller subimages and suitable global methods are adopted to compute the optimal
thresholds for each subimage. It seems that global methods are frequently used for
threshold&g. For this reason, we intend to evaluate some of the global thresholding
methods, namely the Ostu method [27], the Pun method [32], the Johannsen and
Bille Method [15], the method of Kapur et al. [16], the co-occurrence matrix
method [2], the histogram concavity analysis method [35], the method due to Deravi
and Pal [9], the moments preserving method [40], and the m inimum error method
 FIG. 2. (a) Digitized cameraman image; (b) histogram of cameraman image.

[19]. These methods are considered for a comparative study since they are automatic
and in most cases do not require a subjective judgement for finding the best
threshold from a set of optimal threshold values. The above mentioned methods
are applied to three images, “cameraman”, the “building”, and the “model”. The
cameraman image is digitized from the printed version of the same image in [32].
This is done with a raster of 415 X 395 and the quantization is 256 gray levels. The
digitized image and its gray level histogram are shown in Fig. 2. The building image
is also digitized from the printed version of the same image in [32] with a raster of
411 x 403 and quantized to 256 gray levels. The image of the model is digitized
with a raster of 321 X 314 and quantized to 256 levels. The building and the image
of the model are shown in Figs. 3 and 4 along with their gray level histograms. The
threshold values obtained by methods considered above are presented in Table 1.
For the co-occurrence matrix method the distance between cells d was chosen to
be 1 and 5% of the elements of the M matrix that lie farthest from the diagonal was
projected onto the diagonal. Incidentally a threshold value of 63 was found for all
three images. This value did not change when 10% of the elements instead of 5% was
used. However, when d changed from 1 to 2 (resp. 3) the threshold values for the

FIG. 3. (a) Digitized building image; (b) histogram of building image.

 SURVEY OF THRESHOLDING TECHNIQUES 249

FIG. 4. (a) Digitized image of the model; (b) histogram of the model image.

cameraman and the building images did not change, but the threshold value of the
model image changed to 30 (resp. 33). On changing the percentage of elements of M
projected onto the diagonal to 1% and 2% while keeping d values to 2 and 3,
respectively, we have obtained a threshold value 63 in both cases. Thus, whenever
higher values of d are used, it is advisable to decrease the percentage of elements of
A4 that lie farthest from the diagonal and projected onto the diagonal.
The Deravi and Pal method [9] was implemented for the cameraman picture only.
With their method, two local m inima were found, one at 45 and the other at 116.
Since thresholding at 45 does not provide a better binary image we select 116 as the
optimal threshold value. Figures 5-7 show the binary images obtained by the
thresholding methods considered in this section.
7. MEASURES FOR EVALUATING THRESHOLDING METHODS
In digital images, uniformity and shape of the objects play great roles in
separating objects from the background. The amount of agreement of these two
aspects of every binary image with the real image has been evaluated for the three
test images considered in the previous section. The uniformity measure used in this

TABLE 1
Optimal Threshold Value for the Set of Test Images

Optimal threshold value

Method Cameraman Building Model

Co-occurrence matrix 63 63 63
Deravi & Pal 116
Histogram concavity 127 47 127
Johamwn & Bille 75 79 114
Kapur et al. 123 69 81
Minimum error 111 37 113
Moment-preserving 91 76 76
Ostu 86 73 91
b 1321 131 28 63
250 SAHOO ET AL.

FIG. 5. Binary images of the cameraman: (a) co-occurrence matrix method (t’ = 63); (b) Deravi and
Pal method (t* = 116); (c) histogram concavity method (t* = 127); (d) Johannsen and Bille method
(t* = 75); (e) Kapur, Sahoo, and Wong method (t* = 123); (f) minimum error method (t* = 111): (8)
moment-preserving method (z* = 91); (h) Ostu method (r* = 86); (i) Pun method [32] (t* = 131).

paper is adopted from Levine and Nazif [22]. For a given threshold value t, the
uniformity measure U(t) is given by

u1’+ (72”
U(t) = 1 - --y--y

where

Ri = Segmented region i
f (x, y) = The gray level of the pixel (x, y)
c f(w)
(x>Y)ER,
Pi = Ai ’
Ai = Number of pixels in R,, i = 1,2,
 SURVEY OF THRESHOLDING TECHNIQUES 251

FIG. 5.-Continued

and

C = A normalization factor.
Levine and Nazif [22] have included a weighting factor in the calculation of
uniformity measure U for images with more than two regions.
A measure, called shape meusure S, is used for the measurement of the shape of
the object in the test images. The measure S for a given image is calculated as
follows: (a) assign a generalized gradient value A(x, y) to every pixel (x, y); (b) if
the pixel (x, y) has a gray value higher than the average of its neighbors then assign
the “ + ” sign to the generalized gradient value A(x, y), else assign the “ - ” value;
 FIG. 6. Binary images of the building: (a) co-occurrence matrix method (t* = 63); (b) histogram
concavity method (t* = 47); (c) Johannsen and Bille method (t* = 79); (d) Kapur, Sahoo, and Wong
method (t* = 69); (e) minimum error method (f* = 37): (f) moment-preserving method (t* = 76): (g)
Ostu method (t* = 73); (h) Pun method [32] (t* = 28).

(c) compute the shape measure S by using the formula

where frqx, y) is the average gray value in the neighborhood iV(x, y), t is the
threshold value of the image, C is a normalization factor, and

ifx20
W(x) = ( -: ifx < 0.

The computation of the generalized gradient value A(x, y) of the pixel (x, y) is
carried out using the formula

l/2
i
k=l
0; + fi~,(D, + D,>- fiD,(~, - 0,)
1 .

where

D, =f(x+ Ly)-f(x - Lv),

D, = f(x, Y - 1) - f(x> Y + l),
D, =f(x + 1, y + 1) -f(x - 1, y - I),
D4 = f(x + 1, y - 1) - f(x - 1, y + 1).

Using both measures, the threshold value obtained according to each me&i&
mentioned earlier are evaluated. Tables 2a, 2b and 2c show the results of this
 SURVEY OF THRESHOLDING TECHNIQUES 253

FIG. 6.-Continued

evaluation. Note that these values are renormalized according to the best possible
threshold for each measure and the rank of each method is indicated within the
square bracket.
8. DISCUSSION OF EVALUATION RESULTS
Analyzing Table 2a, we find that methods of Ostu, moment-preserving, and
Johannsen and Bille are better threshold selection methods for the cameraman
image with regards to region uniformity and shape measure. For bimodal images
(see Table 3), the optimal values of uniformity and shape measure do not differ
254 SAHOO ET AL.

FIG. 7. Binary images of the model: (a) co-occurrence matrix method (t* = 63); (b) histogram
concavity method (t* = 127); (c) Joharmsen and Bille method (1: = 114); (d) Kapur, Sahoo, and Wong
method (t* = 81); (e) minimum error method (t* = 113); (f) moment-preserving method (I* = 76); (g)
Ostu method (t* = 91).

significantly. Thus for the cameraman image, if a particular method is found to be

good it will also be found equally good with respect to the shape measure.
Therefore, the above mentioned methods in Table 2a have the same rank with
respect to uniformity as well as shape. On the whole, the evaluation results on this
image for the above mentioned methods vary very little.
The building image does not have a distinct birnodal gray level histogram. Thus,
unlike the cameraman image, the Ostu method (see Table 2b) ranks third according
to shape measure and first with respect to region uniformity measure when applied
to the building image. Moment-preserving method ranks second according to
uniformity and ranks first according to shape measure. The performances of
Johannsen and Bille and Kapur et al. methods are next to the Ostu and moment-
preserving method.
Notice that the image of the model does not have a bimodal histogram. For this
image, we find (refer to Table 2c) that Ostu method ranks fhst with respect to
uniformity and ranks fourth with respect to the shape measure. The co-occurrence
matrix method is found to have first rank according to the shape measure and
seventh rank with respect to the uniformity measure. The method due to Kapur et
al. ranks second with respect to uniformity and third according to shape. This result
shows that the Kapur et al. method is fairly consistent with both shape and
uniformity measure for this type of images. The co-occurrence matrix method is
shape oriented and the Ostu method is uniformity oriented. Thus for the image of
the model, which does not have a bimodal histogram, the co-occurrence matrix
 SURVEY OF THRESHOLDING TECHNIQUES 255

FIG. ?‘-Continued

method is found to be a good method if only the shape measure is considered,

whereas the Ostu method is good on the basis of uniformity measure alone.
From this objective evaluation (from Tables 2a, 2b, and 2c), we find the Ostu
method, which is based on discriminant analysis to be one of the better thresholding
methods despite of its many short commings (see [20,34]). Since our measures for
evaluating a thresholding method are based on uniformity and the shape only, it is
not altogether surprising to find Ostu method as one of the better methods. The
moment-preserving method of Tsai is found to be comparable to the Ostu method.
The performance of methods such as Johannsen and Bille as well as Kapur et al. is
almost next to Ostu and the moment-preserving methods.
The thresholded images obtained by various methods (listed in Table 1) reveal
valuable information regarding the thresholding techniques. We incorporate this

TABLE 2a
Evaluation Results for Cameraman Image

Method Threshold Uniformity (U) Shape (S)

Co-occurrence matrix 63 0.7121 [5] 0.7471 [4]

Deravi & Pal 116 0.6190 [6] 0.5740 [6]
Histogram concavity 127 0.3556 [8] 0.3387 [8]
Johannsen & Bille 75 0.9089 [3] 0.9452 [3]
Kapur et al. 123 0.4775 [7] 0.4148 [7]
Minimum error 111 0.7133 [4] 0.6951 [5]
Moment-preserving 91 0.9911 [2] 0.9839 [2]
ostu 86 0.9990 (11 l.oooo [l]
b ~321 131 0.2205 [9] 0.2390 [9]
 SAHOO ET AL.

TABLE 2b
Evaluation Results for Building Image
Method Threshold Uniformity (U) Shape (S)
Co-occurrence matrix 63 0.9098 [5] 0.8863 [5]
Histogram concavity 41 0.5889 [6] 0.5597 [6]
Johannsen & Bille 19 0.9833 [3] 0.9990 [2]
Kapur et al. 69 0.9792 [4] 0.9647 [4]
Minimum error 31 0.3820 [7] 0.2987 [7]
Moment-preserving 16 0.9980 [2] 0.9995 [l]
Ostu 73 0.9986 [l] 0.9931 [3]
Pun [32] 28 0.1988 [8] 0.0131 [X]

TABLE 2c
Evaluation Results for Image of Model

Method Threshold Uniformity (U) Shape (S)

Co-occurrence matrix 63 0.5483 [7] 0.9597 [l]
Histogram concavity 127 0.6246 [6] 0.2775 [7]
Johannsen & Bille 114 0.8334 [5] 0.4543 [6]
Kapur et al. 81 0.9192 [2] 0.7636 [3]
Minimum error 113 0.8465 [3] 0.4682 [5]
Moment-preserving 76 0.8445 [4] 0.7872 [2]
Ostu 91 0.9986 [I] 0.7418 [4]
Pun [32] 63 0.5483 [7] 0.9597 [l]

visual information to supplement our findings. For visual analysis of the binary
images, it seems reasonable to consider some important features such as facial
details and camera of the cameraman picture, edges of the building on the
right-hand side of the building image, and facial and hair features in the model
image. Along with these observations, one should also consider the amount of
distortion and loss of information in the thresholded image. A close look at the
binary images of the cameraman in Fig. 5, we observe that methods such as Pun
(Fig. 5i), histogram concavity (Fig. 5c) and co-occurrence matrix (Fig. 5a) do not
provide a good threshold for this image. This can also be observed from Table 2a.
We also observe that the co-occurrence matrix method yields a binary image in
which the details of the face and the camera of the cameraman are lost, whereas
Pun’s and histogram concavity methods yield binary images with distortions. Also,
Ostu method and moment-preserving method do not retain the details of the face in
the cameraman image. The methods which retain the facial details and other
valuable information of the image are the Kapur et al. method and the m inimum
error method. Notice that the Kapur et aE. method ranks seven (in Table 2a)
according to uniformity as well as shape measure, but it gives a good binary image,
which retains the facial details of the cameraman.
A similar examination of the images in Fig. 6 reveals that the moment-preserving
(Fig. 6f), co-occurrence matrix (Fig. 6a), Ostu (Fig. 6h), Kapur et al. (Fig. 6d),
Johannsen and Bille (Fig. 6c) methods provide reasonably good threshold values for
 SURVEY OF THRESHOLDING TECHNIQUES 257

TABLE 3
Optimal Value for the Uniformity and Shape Measures
-
Measure Cameraman Building Model
-.
Uoiformity 87 74 93
Shape 86 78 57

FIG. 8. Binary images of the cameraman: (a) optimal uniformity (t* = 87); (b) optimal shape
(r* = 86).

the building picture. Among these methods, the Johannsen and Bille method and
the moment-preserving method yield good binary images. Regarding the binary
images in Fig. 7, we observe that the Kapur et al. method and the moment-preserv-
ing method are better threshold selection techniques. For the same image, the Ostu
method also provides a good threshold value. However, it is not as good as the two
methods mentioned.
Thus, from the visual aspect of the binary images as well as the objective
evaluation based on the uniformity and shape measures,we find that Johannsen and
Bille, Kapur et al., moment-preserving, and Ostu methods are good thresholding
methods. In most cases,Pun [32], histogram concavity, and m inimum error methods
do not render good performance in comparison to other methods mentioned in this
section. In an earlier study [12], unsatisfactory performance of the Pun method [32]
was observed by Fernando and Monro.
9. COMMENTS AND CONCLUSION
The indicators used in this paper for evaluating an automatic thresholding
method are shape meaSureand uniformity measure.Table 3 presents the gray levels
which optimize the uniformity and shape measure for the test images. The set of
digital test images, when binarized at the gray levels where S and U attain
maximum, are shown in Figs. 8, 9, and 10. From these binary images it is evident
that one may use these measures for image thresholding.
In summary, we have presented a survey of various thresholding methods and
evaluated the performance of some global methods which are automatic in nature.
All the methods described here optimize some criterion functions and provide
 FIG. 10. Binary images of the model: (a) optimal uniformity (I* = 93); (b) optimal shape (t* = 57).

justifications for such optimization. We only investigate how the optimal values
behave corresponding to the uniformity and shape measures to a given set of test
images. From this study, we conclude that for the above set of test images
Johannsen and Bille, Kapur et al., Tsai’s moment-preserving, and Ostu methods are
reasonably good thresholding methods if one demands more uniformity and better
shape of the object in the binary image. However, whether the aforementioned
methods, in general, will perform well or not remains open.
ACKNOWLEDGMENTS

We are thankful to the referee for helping to improve the presentation of this
paper.

REFERENCES
1. N. Abramson, Information Theon, und Coding, McGraw-Hill, New York, 1963.
2. N. Ahuja and A. Rosenfeld, A note on the use of second-order gray-level statistics for threshold
selection, IEEE Trans. Systems Man Cybernet. SMC-8, 1978, 895-899.
3. M. R. Bartz, Optimizing a video processor for OCR, in Proceedings, Intemarional Joint Conference on
AT, 1969, pp. 79-90.
4. B. Bhanu and 0. Faugeras, Segmentation of images having unimodal distributions, IEEE Trans.
Pultern Anui. Mach. Intell. PAMI-4, 1982, 408-419.
 SURVEY OF THRESHOLDING TECHNIQUES 259

5. S. Boukharouba, J. M. Rebordao, and P. L. Wendel, An amplitude segmentation method based on

the distribution function of an image, Comput. Vision Graphics Image Process. 29, 19S5,47-59.
6. H. Bunke, H. Feistl, H. Neimarm, G. Sagerer, F. Wolf, and G. Zhou, Smoothing, thresholding, and
contour extraction in images from gated blood pool studies, in First IEEE Symp. on Medical
Imaging and Image Interpretation, Berlin 1982.
7. C. K. Chow and T. Kaneko, Boundary detection of radiographic images by a threshold method, in
Proceedings, ZFZP Congress 71, pp. 130-134.
8. C. K. Chow and T. Kaneko, Automatic boundary detection of left ventricle from cineangiograms,
Comput. Biomed. Res. 5, 1972, 338-410.
9. F. Deravi and S. K. Pal, Gray level thresholding using second-order statistics, Pattern Recognit. Zett.
1,1983,417-422.
10. W. Doyle, Operation useful for similarity-invariant pattern recognition, J. Assoc. Comput. Mach. 9,
1962, 259-267.
11. G. Fekete, J. 0. Eklundh, and A. Rosenfeld, Relaxation: Evaluation and applications, IEEE Trans.
Pattern Anal. Mach. Zntell. PAMI-3, 1981, 460-469.
12. S. M. X. Fernando and D. M. Monro, Variable thresholding applied to angiography, in Proceedings,
6th International Conferenceon Pattern Recognition, 1982.
13. S. K. Fu and J. K. Mu, A survey on image segmentation, Pattern Recognit. 13, 1981, 3-16.
14. R. M. Haralick, K. Shanmugam, and I. Dinstein, Texture features for image classification, IEEE
Trans. Systems Man Cybemet. SMC3,1973,610-621.
15. G. Johannsen and J. Bille, A threshold selection method using information measures, in Proceedings,
6th Znt. Conf. Pattern Recognition, Munich, Germany, 1982, pp. 140-143.
16. J. N. Kapur, P. K. Sahoo, and A. K. C. Wong, A new method for gray-level picture thresholding
using the entropy of the histogram, Comput. Vision Graphics Image Process. 29,1985, 273-285.
17. Y. H. Katz, Pattern recognition of meteorological satellite cloud photography, Proceedings, Third
Symp. on Remote Sensing of Environment, 1965, pp 173-214.
18. R. L. Kirby and A. Rosenfeld, A note on the use of (gray level, local average gray level) space as an
aid in thresholding selection, IEEE Trans. Systems Man Cybernet. SMC-9, 1979, 860-864.
19. J. Kittler and J. Illingworth, Minimum error thresholding, Pattern Recognit. 19, 1986,41-47.
20. J. Kittler and J. Illingworth, On threshold selection using clustering criterion, IEEE Trans. Systems
Man Cybernet. SMC-15,1985, 652-655.
21. R. Kohler, A segmentation system based on thresholding, Comput. Graphics Image Process. 15, 1981,
319-338.
22. M. D. Levine and A. M. Nazif, Dynamic measurement of computer generated image segmentations,
IEEE Trans. Pattern Anal. Mach. Zntell. P&H-7,1985, 155-164.
23. D. Mason, I. J. Lauder, D. Rutoritz, and G. Spowart, Measurement of C-Bands in human
chromosomes, Comput. Biol. Med. 5, 1975, 179-201.
24. D. L. M&ram, Region extraction using convergent evidence, Comput. Graphics Image Process. 11,
1979,1-12.
25. T. H. Morrin, A black-white representation of a gray-scale picture, IEEE Trans. Comput. 23, 1974,
184-186.
26. Y. Nakagawa and A. Rosenfeld, Some experiments on variable thresholding, Pattern Recognit. 11,
1979, 191-204.
27. N. Ostu, A threshold selection method from gray-level histogram, IEEE Trans. Systems Man
Cybemet. SMC-8,1978,62-66.
28. D. P. Panda, Segmentation of FLZR Zmages by Pixel Classification, TR-508, University of Maryland
Computer Science Center, 1977.
29. T. Pavlidis, Structural Pattern Recognition, Springer-Verlag, New York, 1977.
30. S. Peleg, A new probabilistic relaxation scheme, IEEE Trans. Pattern Anal. Mach. Zntell. PAMI-2,
1980, 362-369.
31. J. M. S. Prewitt and M. L. Mendelsohn, The analysis of cell images, in Ann. New York Acad. Sci.
Vol128, pp 1035-1053, New York Acad. Sci., New York, 1966.
32. T. Pun, A new method for gray-level picture thresholding using the entropy of the histogram, Signal
Process. 2, 1980, 223-237.
33. T. Pun, Entropic thresholding: A new approach, Comput. Vision Graphics Zmage Process. 16, 1981,
210-239.
34. S. S. Reddi, S. F. Rudin, and H. R Keshavan, An optimal multiple threshold scheme for image
segmentation, IEEE Trans. Systems Man Cybemet. SMC-14,1984, 661-665.
260 SAHOO ET AL

35. A. Rosenfeld and P. De La Terre, Histogram concavity analysis as an aid in threshold selection,
IEEE Trans. Systems Man Cybernet. SMC-13,1983, 231-235.
36. A. Rosenfeld and R. C. Smith, Thresholding using relaxation, IEEE Trans. Pattern Anal. Mach.
Intell. PAMI-3, 1981, 598-606.
37. P. K. Woo, Theory and Applications of Some Measure of Uncertainties, Ph.D. thesis, Department of
Applied Mathematics, University of Waterloo, Waterloo, 1986.
38. R. Southwell, Relaxation Methoak in Engineering Science, A Treatise on Approximate Computatron,
Oxford Univ. Press, London, 1940.
39. R. Southwell, Relaxation Methods in Theoreticai Physics, Oxford Univ. Press (Clarendon), London.
1946.
40. W. Tsai, Moment-preserving thresholding: A new approach, Comput. Vision Graphics Image Process.
29, 1985, 377-393.
41. J. R. Ullman, Binarization using associative addressing, Pattern Recognit. 6, 1974, 127-135.
42. S. Wang and R. M. Haralick, Automatic multithreshold selection, Comput. Vision Graphics Image
Process. 25, 1984, 46-67.
43. S. Watanabe, and the CYBEST group, An automated apparatus for cancer processing CYBEST,
Comput. Vision Graphics Image Process. 3, 1974, 350-358.
44. J. S. We&a, A survey of threshold selection techniques, Comput. Vision Graphics Image Process. 7.
1978, 259-265.
45. J. S. Weszka, R. N. Nagel, and A. Rosenfeld, A threshold selection technique, IEEE Trans. Compuf
C-23, 1974, 1322-1326.
46. J. S. Weszka and A. Rosenfeld, Threshold Selection 4, TR-336, University of Maryland Computer
Science Center, 1974.
47. J. S. Weszka and A. Rosenfeld, Threshold evaluation techniques, IEEE Trans. Systenls Man
Cybernet. SMC-8,1978, 622-629.
48. J. S. Weszka and A. Rosenfeld, Histogram modification for threshold selection, IEEE Trans. Systems
Man Cybernet. SMC-9,1979, 38-51.
49. J. S. Weszka, J. A. Veretuon, and A. Rosenfeld, A Technique for Facilitating Threshold Selection for
Objects Extraction from Digital Pictures, TR-243, University of Maryland Computer Science
Center, 1973.
50. R. N. Wolfe, A dynamic thresholding scheme for quantization of scanned image, in Proceedings,
Automatic Pattern Recognition, 1969, pp. 143-162.
51. A. Y. Wu and A. Rosenfeld, Threshold selection using quadtree, IEEE Trans. Pattern Anal. Mach.
Intell. PAMI-4, 1982, 90-94.
52. S. Zucker, R. Hummel, and A. Rosenfeld, An application of relaxation labelling to line and curve
enhancement, IEEE Trans. Comput. C-26,1977, 394-403.

View publication stats