Int. J. Appl. Math. Comput. Sci., 2010, Vol. 20, No.

2, 317–335
DOI: 10.2478/v10006-010-0024-4

K3M: A UNIVERSAL ALGORITHM FOR IMAGE SKELETONIZATION

AND A REVIEW OF THINNING TECHNIQUES

K HALID SAEED ∗ , M AREK TAB EDZKI

˛ ∗∗
, M ARIUSZ RYBNIK ∗∗∗ , M ARCIN ADAMSKI ∗∗

∗
Faculty of Physics and Applied Computer Science
AGH University of Science and Technology, Al. Mickiewicza 30, 30–059 Cracow, Poland
e-mail: [email protected]
∗∗
Faculty of Computer Science
Białystok Technical University, ul. Wiejska 45 A, 15–351 Białystok, Poland
e-mail: {m.tabedzki,m.adamski}@pb.edu.pl
∗∗∗
Faculty of Mathematics and Informatics
University of Białystok, ul. M. Skłodowskiej-Curie 14, 15–097 Białystok, Poland
e-mail: [email protected]

This paper aims at three aspects closely related to each other: first, it presents the state of the art in the area of thinning
methodologies, by giving descriptions of general ideas of the most significant algorithms with a comparison between
them. Secondly, it proposes a new thinning algorithm that presents interesting properties in terms of processing quality
and algorithm clarity, enriched with examples. Thirdly, the work considers parallelization issues for intrinsically sequential
algorithms of thinning. The main advantage of the suggested algorithm is its universality, which makes it useful and versatile
for a variety of applications.

Keywords: skeletonization, thinning, digital image processing, parallelization, iteration, thinning methodologies, sequen-
tial thinning, parallel thinning.

1. Introduction ve been published on this topic. The thinning algorithms

mentioned in this paper have been selected from among
As a digital image processing technique, thinning has be- the best known published works. These selected works are
en around for many years. Thinning (also termed skele- given here for comparison to show and evaluate the ef-
tonization) is an image pre-processing technique that is forts of other researchers and authors in working out and
important in a number of applications like pattern reco- implementing a number of algorithms and approaches for
gnition, data compression and data storage (Gonzalez and a variety of applications.
Woods, 2007). The history of thinning covers over fifty
years. In the 1950s Dinnen (1955) found that an averaging The first trials conducted in this field by the au-
operation over a square window with a high threshold re- thors of this paper date back to 1999 (Saeed and Nie-
sulted in a thinning of the input image. Then Kirsch et al. dzielski, 1999). This was to aid recognition algorithms of
(1958) mentioned that the “custer” operation was an early cursive-character letters, which required thinning before
attempt to obtain a thinline representation of certain cha- classification. This work is an attempt to propose a new
racter patterns. However, still there is no ultimate solution criterion and employ it to various applications. The pre-
suitable for each individual area of application (Jaisimha sented method leads to a fast recognition algorithm of both
et al., 1994; Jang and Chin, 1992; Lam et al., 1992; Lee typewritten and handwritten texts with the possibility of
et al., 1991). This is due to the complicated nature of applying it to thin images of a variety of objects. The ba-
the processing as well as various evaluation measures that sic idea of thinning according to this approach lies in the
should be considered. generation of a skeleton with a one-pixel width. Howe-
There are more than one thousand algorithms that ha- ver, to serve more complicated cases, some modifications

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 318 K. Saeed et al.



 $

%& 


# $ '%




  (%



)%



•


 
•     
• 
 

  
•        
•  
   
•  
 

•   ! " 


Fig. 1. Thinning algorithms taxonomy using two criteria: processing and methodology.

have been made to the whole criterion. The method sugge- This survey is mainly based on algorithms mostly
sted by the authors is universal and simple. It consists of known to the authors, available either as papers in inter-
several fast and easy-to-implement deleting phases. The national conference proceedings, periodical and journal
obtained experimental results have also shown a success- transactions or book chapters. The central role in the su-
ful application of the method in image pre-processing as rvey was played by Pavlidis (1982a), Lam et al. (1992),
well as image preparation for feature extraction and clas- Malina et al. (2002) and, finally, Klette and Rosenfeld
sification. The method proved its universality—not only (2004). This is in addition to lots of individual works sho-
does it work for printed words, but it has also shown a wing different methods and algorithms of thinning.
remarkable success rate on thinning handwritten scripts, Generally, each of the presented algorithms has its
signatures, numbers, characters of non-Latin-based alpha- own advantages and disadvantages, and each has its appli-
bets (Arabic, Japanese) and even biometric images such cations where it performs better than others. Therefore, it
as faces or fingerprints. The paper is organized as follows. is often difficult to directly compare the results.
Section 2 presents the state of the art in the following or-
der: the proposed taxonomy of thinning algorithms, a de-
scription of general ideas of the most significant ones and 2.1. Taxonomy. All thinning algorithms can be clas-
a short comparison in terms of efficiency. Section 3 pre- sified as either iterative or non-iterative. In iterative me-
sents the thinning algorithm proposed by the authors in thods, thinning algorithms produce a skeleton by exami-
detail. Section 4 presents the results of the K3M algorithm ning and deleting contour pixels through an iterative pro-
and the comparison with other algorithms. Section 5 de- cess in either sequential or parallel way.
scribes the concepts for the parallelization of intrinsically Parallel algorithms may also be further classified ac-
sequential thinning algorithms. Finally, Section 6 conclu- cording to their performance, i.e., in 4-, 2-, or 1-subcycle
des the work and shows perspectives for future develop- manners. The latter (1-subcycle parallel algorithms) ha-
ment. ve always received more considerable attention in the re-
search area of parallel thinning as they have reduced the
computation time in a number of iterations, and that is
2. State of the art
why they are sometimes called one-pass or fully parallel
The survey presented in this section introduces a historical algorithms (Chen and Hsu, 1989b; Chen, 1996; Guo and
review covering over 40 years of works on thinning algo- Hall, 1992).
rithms, together with their basic principles and the state of In sequential thinning algorithms, contour points are
the art. examined for deletion in a predetermined order, and this

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 319

can be accomplished by either raster scanning or follo- proach, in which he defined the ‘crossing number’—the
wing the image by contour pixels. In parallel thinning al- number of times one crosses over from a white point to
gorithms, pixels are examined for deletion on the basis of a black point when the eight neighbours of a tested pixel
results obtained only from the previous iteration. That is are traversed in sequence. The image is scanned from left
why parallel thinning algorithms are suitable for imple- to right and from top to bottom, and pixels are marked
mentation in parallel processors. for deletion under additional conditions to prevent ero-
Non-iterative (non-pixel based) thinning algorithms sion, maintain connectivity and preserve two-pixel wide
produce a certain median or centre line of the pattern to lines. This method was also used with gray-scale images.
be thinned directly in one pass, without examining all the Rosenfeld (1975) showed many new aspects and so-
individual pixels. lved serious problems commonly met in thinning algori-
The methods used in the works presented in this sec- thms. He was the first to evaluate the necessary and suf-
tion may also be categorized as one of the following me- ficient conditions for preserving topology while deleting
thodologies (Fig. 1). border points in parallel process. Thereby, he solved the
problem of erosion in the thinning of diagonal lines. In his
2.2. General ideas and examples. In this section a de- algorithms, he considered a 3×3 local neighbourhood. He
tailed description of most of the basic algorithms of thin- proved that ‘a border pixel is removed only if its neighbo-
ning is presented in order to show how others attack the urhood has only one black component (i.e., the Hilditch
problem of skeletonization. Some of them are explained crossing number is 1) and has at least two black neighbo-
in detail in (Saeed, 2004). urs.’
Rutovitz (1966) proposed a parallel algorithm that Arcelli’s parallel behaviour algorithm (Arcelli and
formed a basis for many thinning algorithms. Image pixels di Baja, 1978) uses two 3 × 3 thinning windows toge-
are examined for deletion in an iterative process (not divi- ther with their 90◦ rotations as masks for pixel deletion.
ded into subcycles). Contour pixels with at least two black It removes pixels from eight borders in the following or-
neighbours, and where all its neighbours are 4-connected der: north-west, west, and so forth. However, it does not
(sticking together with its edge), are marked for deletion. remove all deletable pixels.
This may result in excessive erosion, it does not reduce Dyer and Rosenfeld (1979) introduced the algorithm
diagonal lines to a unit pixel width and the skeleton is not for thinning gray-scale pictures. It is based on a generali-
located centrally (due to the asymmetric nature of its con- zed concept of pixel connectivity: two pixels are “connec-
ditions). ted” if there is a path joining them with no pixel lighter
Blum (1967) presented skeletonization as transfor- than either of them. Using this idea, the thinned version
mation for the aim of shape description. He called this on an image can be obtained by changing each pixel gray
transformation MAT—Medial Axis Transformation. Figu- level to the minimum of its neighbor’s gray levels.
re 2 shows examples of obtaining skeletons from some Pavlidis (1980; 1981; 1982a; 1982b) introduced the
objects by this method. definition of “multiple pixels” for the first time—points
that are traversed more than once during contour tracing,
points with no neighbours in the interior and points on
two-pixel-wide lines. If only the “multiple pixels” from
every tracing are retained, the result may not be a connec-
ted skeleton. Therefore, “multiple pixels” are called ske-
letal, as well as eight neighbours of skeletal pixels from
a previous iteration. This does not result in a one-pixel-
wide skeleton. Contour pixels are traced sequentially.
Then, Pavlidis (1982a) proposed a combination of
parallel and sequential operations—a pattern could be di-
vided into fields, which are processed independently by
parallel processors, each operating sequentially on its own
segment.
Arcelli’s algorithm (Arcelli, 1981) is of a sequential
type. It uses contour tracing to find the pixels for deletion.
This approach allows reducing the computational time ne-
eded to obtain the skeleton. Contour analysis is performed
Fig. 2. Illustration of the result of thinning shapes by Blum’s to find regions to be represented by the skeleton branches.
algorithm. The method of Arcelli and di Baja (1981) is also a se-
quential algorithm—successive iterations consider the re-
Hilditch (1968; 1969) worked out a sequential ap- moval of contour elements. Regions that can be regarded

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 320 K. Saeed et al.

as significant protrusions are detected before applying the geometric features of the image (see Fig. 3(b)).
removal operations. In (Arcelli and di Baja, 1987), the au- Chen and Hsu worked out an interesting algorithm
thors used the Pavlidis concept of ‘multiple pixels’ (pixels modifying it through a series of papers (1989b; 1989a;
are placed where the contour self-interacts). The paper gi- 1990; 1988). Chen (1996) reached a 1-subcycle parallel
ves a necessary definition to satisfactorily detect such pi- thinning algorithm producing a one-pixel-wide skeleton
xels. Then Arcelli and di Baja (1989) developed a sequ- that preserves significant geometric features of patterns
ential algorithm that used a 4-distance transform to find (Fig. 3(c)). He developed a method embedded in a pa-
a set of skeletal pixels (two-pixel wide, at most) within rallel algorithm (Chen and Hsu, 1989b) to produce bias-
one raster scan of the image. Then, another inspection of reduced skeletons using, as a major factor, what he defined
the picture removes the unnecessary pixels. the HDP—Hidden Deletable Pixel detected by algorithms
The algorithm of Favre and Keller (1983) covers one- of vector analysis. The HDP is detected and removed, so
subcycle parallel thinning using syntactic rules and a ca- that the bias skeleton can be reduced. Despite the satis-
scade of tables to determine a new pixel code, on the factory results of thinning, Chen’s algorithm did not show
basis of the distance to the background. Two-scan post- any significant application to object-images, handwritten
processing is applied to obtain the final skeleton. words or at least machine-written alphabets of a cursive
Ammann and Sartori-Angus (1985) proposed a fast character. The results of thinning curved parts of machine-
thinning algorithm where the image is compressed to a re- written English letters, for example, are not as good as
duced gray-scale representation. This is then thresholded those of straight lines. This is demonstrated by the thin-
(with connectivity constraints) to a binary image to which ning results of this algorithm in Fig. 3(c).
an existing thinning algorithm (Arcelli and di Baja, 1981)
was applied. The resulting skeleton is then expanded to its
original scale.
Chin et al. (1987) wrote a paper presenting a one-
subcycle thinning algorithm and its parallel implementa-
tion. They called it a ‘one-pass algorithm’. They paid at-
tention to the idea of bias skeletons, the basic approach
used later by Chen (Chen and Hsu, 1989b; Chen, 1996) in
his modified algorithm. Bias skeletons appear in the junc-
tion of lines that form angles less than 90◦ (Fig. 3(a)).
They mentioned two reasons behind generating bias ske-
letons. One is that the restoring templates are of an even
size and hence the pixel removal process is not symme-
trical, unless we are dealing with systematic design, as re-
storing templates are only used in maintaining the connec-
tivity of the final skeleton. The other factor is due to the (a) (b) (c)
side effect of the thinning templates, that is, they remove
pixels from convex corners faster than from concave cor- Fig. 3. Thinning by CWSI (Chin et al., 1987) (a), AFP3 (Guo
ners. A sample of thinning results using this algorithm is and Hall, 1992) (b), CYS (Chen, 1996) (c).
given in Fig. 3(a).
Baruch (1988) presented a noniterative, non-pixel ba- Parker et al. (1994) introduced the force-based ap-
sed algorithm, which produces a skeleton in one pass, by proach with a new idea for thinning strategy based on a de-
line-following. The line is followed by a window of va- finition of a ‘skeletal pixel’ as being as far from the object
riable size, the skeleton is the line connecting centres of outline as possible while maintaining basic connectivity
successive windows. The method is less sensitive to noise properties. Accordingly, the basic idea is that a skeleton is
than conventional thinning algorithms. a global property of a binary object, and that the boundary
The algorithm of Guo and Hall (1989; 1992) is ano- should be used to locate the skeleton pixels. The criterion
ther example of parallel thinning. The contour pixels are is summarized as follows: the background pixels which
examined for deletion in an iterative process. The decision are adjacent to the boundary act as if they exerted a for-
is based on a 3 × 3 neighbourhood. Each iteration is di- ce; the skeletal pixels lie in areas having the ridges of this
vided into two subcycles—one deleting the north and east force field and are located by searching where directions
pixels and the other for south and west. This algorithm do- of the force vectors change significantly. Figure 4 shows
es not give a one-pixel-wide skeleton. Although it shows how to compute the force at a pixel location and the vector
very good results in thinning and provides a rather good sum of forces.
skeleton, the algorithm and its 1-subcycle modified ver- Zhang and Wang (1996) created the 2-Subiteration
sion (Guo and Hall, 1992) do not preserve all significant Parallel-Thinning Algorithm with Templates—PTA2T.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 321
Original Altuwaijri Zhang’s Original Altuwaijri Zhang’s
Image Thinning Thinning Image Thinning Thinning
2

1
1
2

Pixels exerting a force

‘Invisible’ pixels
3

Object pixels

Pixel under consideration

(a) (b)
Fig. 6. Altuwaijri’s thinning of images with added noise com-
Fig. 4. Force-based thinning: computing the force acting at a gi- pared with thinning by Zhang’s algorithm.
ven pixel (a), vector sum of forces (b).

character alphabets such a problem is serious as the re-

This algorithm actually represents a modification to Guo sulting letter may resemble another letter, rather than the
and Hall’s approach. In an iterative process the deletion correct one. Figure 7 shows how the redundant loop may
decision is based on eleven 3×3 thinning templates. There change the letter Seen into the letter Ssad.
are two subcycles—the second uses templates rotated by
180 degrees. The algorithm preserves the connectivity of
patterns and produces one-pixel-wide skeletons (Fig. 5).
In the paper, Zhang claimed this algorithm was faster than
Guo and Hall’s.

Fig. 7. Redundant loops may transform the script Seen into ano-
(a) (b) (c) ther one—Ssad.

Fig. 5. Thinning results of the image in (a) using Guo and Hall Andreadis et al. (2000) proposed a method for obta-
(b) and Zhang (c) methods. Empty rectangles indicate ining a skeleton from a color image using morphological
the places of the removed pixels operations on vectors defined in the HSV color space. In
the presented method the extraction of a skeleton is pre-
The Altuwaijri algorithm—ART2 NN (Altuwaijri and ceded by the ordering of vectors based on their individual
Bayoumi, 1998) is a data-image-clustering based thinning components (hue, saturation and value).
algorithm that employs neural networks and adaptive reso- KMM (Saeed and Niedzielski, 1999; Saeed, 2001;
nance theory for image clustering. The image is an Arabic Saeed et al., 2001) is the authors’ original contribution
character (Altuwaijri and Bayoumi, 1998). Altuwaijri cla- based on an original approach described in (Saeed and
ims the algorithm produces skeletons which are superior Niedzielski, 1999) and modified in a series of works
to the outputs of the conventional algorithms. He com- (Saeed, 2001; Saeed et al., 2001). The algorithms acro-
pared it (Fig. 6) with Zhang’s algorithm given in (Zhang nym comes from the initials of the authors: Khalid, Marek
and Wang, 1996). ART2 showed higher data-reduction ef- and Mariusz. Basically, the algorithm is a sequential itera-
ficiency and much simpler skeletons with less noise spurs tive one. It produces a one-pixel-wide continuous skeleton
and reduced time complexity. from an image. The details of the thinning algorithm are
The main advantage of this method, however, is its given in (Saeed et al., 2001). The main goal of the algori-
ability to remove the redundant loops appearing in the fi- thm was
nal thinned shape in many thinning algorithms. This pro- 1. to obtain a continuous line of symmetry (the name
blem may not seem particularly important for the prese- given by Blum to the skeleton (Blum, 1967)), and
rvation of essential geometric features in most alphabe-
tical scripts. However, in Arabic and many other cursive 2. to thin all kinds of images to their skeleton—digits,

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 322 K. Saeed et al.

alphabet letters, signatures, pictures and medical 4 × 3 windows. In order to reduce the loss of informa-
images in their two-dimensional views. tion when thinning shapes like in Fig. 9(a), the authors
introduce a parameter R,
The ideal skeleton was not the aim of the criterion. In
fact, there is no algorithm that yields an ideal skeleton in A(s1 )
R= , (1)
any case. The most important thing is usually the retaining A(s2 )
of pixels that preserve the significant geometric features
of the given pattern in all kinds of images. Figure 8 shows where A(s) is the area of s pixels, s1 is the image skeleton,
KMM thinning results when applied to the same figures s2 is the image contour.
given in (Malina et al., 2002) and shown earlier in Fig. 2. Based on its value, the algorithm returns either a thin-
Notice that the two circles differ from each other by one ned image (Fig. 9(b)) or its contour (Fig. 9(c)). If R is less
pixel-length in their diameters. The circle of a diameter
with an even number of pixels is thinned to a dot, while
that of a diameter with an odd number of pixels is thinned
to an angular shape. Note also that a rectangle is thinned
to a line and seems to be more appropriate than Blum’s
skeleton (Blum, 1967).
(a) (b) (c)
KMM was under continuous modification. Its latest
version for universal use will be given in Section 3 as Fig. 9. Image (a), its thinned shape (b) and its contour (c).
K3M—Khalid, Marek, Mariusz, Marcin.
than the threshold, the image contour is retained. The thre-
shold value used by the authors was 0.4.
Ji and Feng (2004) created a method that interprets
the image as a 2D thermal conductor that consists of pi-
xels, where pixel intensity represents the temperature. Pi-
xels with a higher intensity value have higher temperature,
while the background pixels have the lowest temperature.
The thinning task is considered as an inverse process of
heat conduction. The thinning process follows the inverse
of the temperature flow, from cooler (background pixels)
to warmer areas (object pixels with higher intensity), as in
Fig. 10.

Fig. 8. Figures presented in (Malina et al., 2002) thinned with

KMM (Saeed, 2004) (compare with Fig. 2).

The algorithm presented by Ahmed and Ward (2002)

is an example of parallel thinning with the decision of de-
letion made upon an algorithm based on quantity and po- (a) (b)
sition of neighbours of each pixel. With 20 a rules, a pixel
is classified into one of four classes, and then a decision is Fig. 10. Original image (a) and its heat-flow direction map (b).
made. Two-pixel-wide lines are treated separately.
Huang et al. (2003) proposed another parallel thin- The first step of the algorithm is the calculation of the
ning algorithm. Pixel elimination rules are based on 3 × 3 Heat-Flow Direction Map (HFDM), which gives the di-
windows considering all kinds of relations (256) formed rections and magnitude of the time-inversed heat flow for
by 8-neighbours of the object pixel. All the rules are ap- each point of the image. Then for each pixel of the image
plied simultaneously to each pixel. In order to keep the the intensity is subtracted and added to the pixel selected
connectivity of even pixel-width lines, additional checks based on information provided by the HFDM. The latter
are applied before deleting the pixels. These additional step is iterated until a completely thinned image is obta-
checks involve using rules based on 4 × 4, 3 × 4 and ined. The disadvantages of the proposed method lie in the

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 323

fact that it does not guarantee a one-pixel width skeleton

and that the computation of the right number of iterations
to preserve the shape topology is a difficult task.
Quadros et al. (2004) proposed an algorithm to ge-
nerate a disconnected, three-dimensional (3D) skeleton.
A discrete skeleton is generated by propagating a wave
from the boundary towards the interior on an octree latti- (a) (b)
ce of an input solid model. As the wave propagates, the
distance from the boundary and the direction of the wa-
ve front are calculated at the lattice-nodes (vertices) of the
new front. An example of their results is shown in Fig. 11.

(c) (d)

Fig. 12. Pixel numbering convention (a), example of pixel con-

figuration (b), resulting graph (c), and adjacency matrix
(d).

(a) (b)

Fig. 11. Example of a 3D skeleton (Quadros et al., 2004) obta-

ined by wave propagation: graphics facets of input (a),
trimmed discrete 3D skeleton (b).
(a) (b)
Rockett’s algorithm (Rockett, 2005) aimed at impro-
Fig. 13. Example of a skeleton obtained using Ahmed and
ving the rule based algorithm presented by Ahmed and Ward’s method (a), compared with the result of Roc-
Ward (2002). They observed the number of cases whe- kett’s algorithm (b).
re the A-W algorithm failed to produce a centre line of a
single pixel width. To eliminate such cases they adopted
a two-stage thinning procedure which used the A-W ru-
les to thin down to a skeleton which includes 2-pixel wide
lines. As a second stage, they examined the 2-pixel wide
lines in the skeleton produced by the first processing sta-
ge to see which pixels can be deleted without compromi- Fig. 14. Examples of greyscale images of handwritten letters
sing the connectivity of the skeleton. In order to determi- and their skeletons extracted by means of the method
ne if the deletion of a pixel in a 2-pixel wide line disrupts presented in (Pervouchine and Leedham, 2005).
connectivity, they built an adjacency matrix from an undi-
rected graph of the local pixel connectivity over its eight
neighbours (Fig. 12). thod creates a family of skeletons parameterized by two
To retain connectivity, before deleting the pixel the user-specified numbers that determine respectively the si-
algorithm checks if every row (or column) of the adjacen- ze of the curve and surface features on the skeleton.
cy matrix contains at least one non-zero entry. An example You and Tang (2007) extracted the skeleton of a cha-
of Rockett’s results is presented in Fig. 13. racter on the basis of the wavelet transform (Mallat, 1989).
Pervouchine and Leedham (2005) constructed the The skeletonization algorithm starts with the extraction of
skeleton in three steps directly from the greyscale ima- a primary skeleton in a regular region followed by amend-
ge and represent it as a set of curves which in turn are ment processing of the primary skeletons and connecting
represented as cubic B-splines. The method was designed them in the singular region (Fig. 15).
to extract the skeleton which is very close to human per- Wan et al. (2008) proposed a three-stage skeletoni-
ception of the original pen tip trajectory (Fig. 14). zation algorithm for shaped fiber recognition. In the first
Ju et al. (2007) proposed a method for computing stage, the Euclidean distance transform is applied to the
skeletons of volumetric models by alternating thinning fiber image. In the second stage, the local maximal disc
and a novel skeleton pruning routine. The presented me- centres are determined as the true skeleton points, regar-

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 324 K. Saeed et al.

will preserve the structure of the image very well. As

a result, after the transformation, lines that contribute
to the letter are of a shape similar to the original and
cross in a similar way.

2. The input image is binary, where each pixel is either

an image pixel (encoded as 1) or a background pi-
xel (encoded as 0). Therefore, colour and grey-scaled
images are transformed to binary form.

3. The K3M algorithm is a sequential iterative algori-

thm: the iterations are repeated in sequence until no
(a) (b) modification is made to the image during the whole
Fig. 15. Examples of images (a) and their thinned versions (b) iteration.
using the algorithm described in (You and Tang, 2007).
4. Each iteration involves seven phases (numbered
from 0 to 6). The number of phases is determined
dless of their connectivity. In the last step, a full skeleton is in an empirical way.
generated by linking isolated skeleton points, and pruning
spurs arising from edge noise (Fig. 16). 5. Pixel neighbours are image pixels that are in the close
vicinity of the tested pixel.

6. Border pixels are those image pixels that stick to the

background (they have no more than seven neighbo-
urs). Therefore, they are considered potential candi-
dates for deletion.

7. Thinning decisions are made for border pixels, based

on 3 × 3 neighbourhood templates of the tested pi-
xel. The 3 × 3 neighbourhood was chosen because of
its large size capable to provide satisfactory thinning
results. Larger neighbourhoods, however, would de-
mand more computational complexity and execution
time.

8. The thinning decision is either to delete the pixel

(a) (b) (change an image pixel into a background one) or
to retain it (the image pixel is not changed). The
Fig. 16. Examples of star and trilobal shaped fibres (a) and their possible neighbourhood configurations are encoded
skeletons (b) obtained using the method presented in using values in the range [0, 255] named neighbour
(Wan et al., 2008). weights. The encoded neighbourhood configurations
represented by neighbour weights are individual for
Table 1 concludes the presentation of selected works each phase and are stored in lookup arrays.
by giving a summary of their most important features.
9. After the iterations are finished, a supplemental pro-
3. K3M: A modified KMM algorithm cedure is required to produce a one-pixel width ske-
leton. This is because the iterative part of the algori-
This section presents the K3M (improved and generalized thm is aimed at producing a fine skeleton and hence
KMM) thinning algorithm in detail. First, definitions and the algorithm is unable to produce a one-pixel width
assumptions are presented, and then a detailed algorithm skeleton.
description is introduced with flowcharts and neighbour-
hood lookup arrays.
3.2. Algorithm description. K3M consists of an ite-
rative part of seven phases and a single additional phase
3.1. Assumptions and definitions.
at the end, which is responsible for producing a one pixel-
1. According to K3M, the aim of thinning is defined as width skeleton. The overview of the algorithm is presented
the reduction of the number of pixels of an image that in Fig. 17.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 325

Table 1. Basic algorithms of thinning.

Year Author Type Remarks
1966 Rutovitz (1966) parallel, iterative, not divided can result in excessive erosion, does not reduce diagonal
into subcycles lines to a unit pixel width, the skeleton does not lie cen-
trally
1967 Blum (1967) medial axis transformation, no-
niterative
1969 Hilditch (1969) sequential maintain connectivity, preserve two-pixel wide lines
1975 Rosenfeld (1975) parallel solves the problem of erosion in the thinning of diagonal
lines
1978 Arcelli and di Baja (1978) parallel does not remove all deletable pixels
1979 Dyer and Rosenfeld (1979) interative directly thins gray-scale images
1980 Pavlidis (1980) sequential, combined sequen- does not result in a one-pixel-wide skeleton
tial and parallel
1981 Arcelli (1981) sequential uses contour tracing to reduce computational time
1983 Favre and Keller (1983) one-subcycle parallel
1985 Ammann and Sartori-Angus (1985) sequential uses image compression before thinning to reduce compu-
tational time
1987 Chin et al. (1987) 1-subcycle, parallel skeleton biased in line junctions
1988 Baruch (1988) noniterative produces skeleton in one pass, by line following
1989 Guo and Hall (1989) iterative, 2-subcycles does not give a one-pixel-wide skeleton
1989 Chen and Hsu (1989a; 1989b) 1-subcycle, parallel bias-reduced skeletons
1994 Parker et al. (1994) force-based acquires a skeleton by using force vectors
1996 Zhang and Wang (1996) 2-subcycles, parallel preserves connectivity, produces one-pixel-wide skeletons
1998 Altuwaijri and Bayoumi (1998) neural networks based redundant loops removal—disadvantage in certain alpha-
bets
2000 Andreadis et al. (2000) interative obtains skeletons from color images
2001 KMM (Saeed et al., 2001) sequential, iterative maintain connectivity, bias-reduced skeletons
2002 Ahmed and Ward (2002) parallel fails to produce a 1-pixel width line in some patterns
2003 Huang et al. (2003) parallel
2004 Ji and Feng (2004) noniterative does not guarantees one-pixel width skeleton, problems
with choosing the right number of iterations
2004 Quadros et al. (2004) 3-dimensional uses wave propagating from the boundary towards the in-
terior
2005 Rockett (2005) parallel produces one-pixel-wide skeletons, uses a two-stage thin-
ning procedure
2005 Pervouchine and Leeedham (2005) gray-scale image thinning represents a skeleton as cubic B-splines
2007 Ju et al. (2007) creates a family of skeletons with various sizes of the cu-
rve and surface features
2007 You and Tang (2007) wavelet based extracted skeleton is centered inside the underlying stroke
2008 Wan et al. (2008) distance map based three-stage skeletonization algorithm for shaped fiber re-
cognition

Iterative part overview • Decision: If any modification was made during the
current iteration, return to Phase 0.
• Phase 0: Mark borders for examination.

• Phase 1: Delete the borders that have 3 neighbours Thinning to a one-pixel width skeleton. This phase
sticking each other. aims at producing skeletons consisting of pixels with only
two neighbours unless located at a junction. This is a si-
• Phase 2: Delete the borders that have 3 or 4 neighbo- gnificant advantage of a one-pixel width skeleton as it po-
urs sticking each other. tentially allows producing structural graphs of the exami-
ned image.
• Phase 3: Delete the borders that have 3, 4 or 5 neigh- The essential property of each pixel, which is used
bours sticking each other. for quick determination of its neighbourhood configura-
tion, is the neighbour weight. It is an 8-bit number, where
• Phase 4: Delete the borders that have 3, 4, 5 or 6
subsequent bit values correspond to neighbours, starting
neighbours sticking each other.
from the pixel above and going clockwise. It is calculated
• Phase 5: Delete the borders that have 3, 4, 5, 6 or 7 with the use of the neighbourhood bit values matrix (3),
neighbours sticking each other. as in (2), where w(x, y) is the neighbour weight of pixel
(x, y) and img(x, y) is the binary value of the image pixel
• Phase 6: Unmark the remaining borders. at coordinates (x, y).

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 326 K. Saeed et al.







 







 












!

  

"  

#  




$% % 

 




 

 





 




Fig. 17. K3M algorithm simplified flowchart.


Neighbour weight calculation

Fig. 18. Flowchart of phase 0—marking borders.
1
 1

w(x, y) = N (i + 1, j + 1) · img(x + i, y + j).
i=−1 j=−1 They serve for marking borders and determining thinning
(2) decisions. The components of arrays Ai (Phases 1 to 5)
follow the rules mentioned in the algorithm description
Neighbourhood bit values matrix concerning the number of neighbours. The contents of ar-
⎡ ⎤ ray A1pix (thinning to a one-pixel width skeleton) was de-
128 1 2 termined on the basis of several experiments and the au-
N = ⎣ 64 0 4 ⎦ . (3) thors’ original KMM algorithm (Saeed et al., 2001).
32 16 8
Figure 18 presents the flowchart of Phase 0, aimed at List of neighbourhood lookup arrays
marking borders (pixels that are candidates for deletion).
The iterative phases 1 to 5 are very similar to each • A0 : lookup array for Phase 0—marking borders,
other and aim at deleting pixels with a growing number of
• A1 : lookup array for Phase 1—deleting pixels having
sticking neighbours. The phases may be presented using
3 sticking neighbours,
a common flowchart with parameter i that changes from
1 to 5 (see Fig. 19) and determines the employed lookup • A2 : lookup array for Phase 2—deleting pixels having
array Ai . 3 or 4 sticking neighbours,

3.3. Neighbourhood lookup arrays. In this section • A3 : lookup array for Phase 3—deleting pixels having
some specific neighbourhood lookup arrays are given. 3, 4 or 5 sticking neighbours,

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 327

143, 159, 191, 192, 193, 195, 199, 207, 223, 224,
225, 227, 231, 239, 240, 241, 243, 247, 248, 249,
251, 252, 253, 254},



• A1 = {7, 14, 28, 56, 112, 131, 193, 224},
• A2 = {7, 14, 15, 28, 30, 56, 60, 112, 120, 131, 135,

 
 193, 195, 224, 225, 240},
• A3 = {7, 14, 15, 28, 30, 31, 56, 60, 62, 112, 120,
124, 131, 135, 143, 193, 195, 199, 224, 225, 227,
240, 241, 248},

 

 • A4 = {7, 14, 15, 28, 30, 31, 56, 60, 62, 63, 112, 120,


 124, 126, 131, 135, 143, 159, 193, 195, 199, 207,
224, 225, 227, 231, 240, 241, 243, 248, 249, 252},
• A5 = {7, 14, 15, 28, 30, 31, 56, 60, 62, 63, 112, 120,
124, 126, 131, 135, 143, 159, 191, 193, 195, 199,
207, 224, 225, 227, 231, 239, 240, 241, 243, 248,

   249, 251, 252, 254},
  • A1pix = {3, 6, 7, 12, 14, 15, 24, 28, 30, 31, 48, 56,

 60, 62, 63, 96, 112, 120, 124, 126, 127, 129, 131,
135, 143, 159, 191, 192, 193, 195, 199, 207, 223,
224, 225, 227, 231, 239, 240, 241, 243, 247, 248,

 249, 251, 252, 253, 254}.

 
 
  3.4. Modifications to KMM. K3M may be seen as
an evolved KMM algorithm as both of them are based
on iterative sequential thinning. K3M is more systematic.
KMM, however, was essentially based on a single lookup
array while K3M uses many of them in an understandable


  pattern. Therefore, K3M involves more iterative phases,

  namely, six, in comparison with KMM, which contains
only four. The main motivation to develop a new K3M
algorithm comes from the fact that the previous attempt
(KMM) was to develop a more systematic approach to


image thinning. As was verified by experiments, the re-
sults are subjectively better so the attempt has proved its
 success rate.

Fig. 19. Flowchart of Phase i (i ∈ {1, 2, 3, 4, 5}) deleting pi- 4. K3M results and comparison
xels. In this section the most interesting thinning results of the
K3M algorithm are presented and discussed. We will the-
• A4 : lookup array for Phase 4—deleting pixels having refore test the algorithm on a variety of inputs—isolated
3, 4, 5 or 6 sticking neighbours, letters, handwritten words and, finally, graphics and sym-
bols, regarding skeleton continuity, shape preservation
• A5 : lookup array for Phase 5—deleting pixels having and a 1-pixel-wide skeleton. The goal is to verify if the
3, 4, 5, 6 or 7 sticking neighbours, performance complies with the expectations.
• A1pix : lookup array used for thinning to a one-pixel
width skeleton. 4.1. Isolated letters. As can be observed in Fig. 20,
K3M preserves the letter shapes. For numerous applica-
tions it is especially important and essential to maintain
Components of neighbourhood lookup arrays
the original connectivity of the script lines as well as the
• A0 = {3, 6, 7, 12, 14, 15, 24, 28, 30, 31, 48, 56, 60, straight angles. K3M does an excellent job here, as can be
62, 63, 96, 112, 120, 124, 126, 127, 129, 131, 135, seen in the letter ‘B’, for example.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 328 K. Saeed et al.

loops letters, like the second ‘e’ in the word ‘seven’, it

is obvious that some information is lost. It is impossible
to avoid such problems without additional steps to create
a real loop from such a letter. Irregular dots are transferred
to short lines rather than to real dots, which is a common
trait for most thinning algorithms.

4.3. Graphical symbols. In thinning graphical sym-

bols (Fig. 22) and graphic-alike alphabets (Fig. 23), it is
especially important to keep the right angles and intercon-
nections close in terms of the shape to the original pictu-
re. One can see that these aspects are well managed by the
K3M algorithm. However, occasional deformations of ori-
ginally straight diagonally placed lines can be observed.

(a) (b)

Fig. 22. Thinning graphical symbols with K3M: original ima-

ges (a), their thinned shapes (b).

Fig. 20. Thinning isolated letters with K3M.

4.2. Handwritten words. In most cases printed words

are thinned (with some exception in the case of bad quality
texts) in the same way as several isolated letters. Handw-
ritten words, however, are interesting as they contain one
or more connections between individual letters, extra lo-
ops, and irregularities, which is worth studying.
Fig. 23. Results of thinning a Japanese symbol with K3M.

4.4. Comparison of results between K3M and other

algorithms. Table 2 shows the results of thinning when
using the K3M approach in comparison with some of the
above-mentioned algorithms.
Table 3, however, presents a comparison between
K3M and other algorithms considering several important
aspects, like the output width, angles preservation and to-
pology preservation.
Fig. 21. Handwritten words thinned with K3M.

As seen in Fig. 21, the overall shapes of the examined 4.5. Advantages and drawbacks of K3M. K3M has
words are well preserved. However, for possibly closed- the following feasible advantages:

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 329

Table 2. Results of thinning by Guo and Hall’s, Zhang’s, A-W, KMM and K3M algorithms.
Image

Method

Guo and Hall’s (Guo and Hall, 1992)

Zhang’s (Zhang and Wang, 1996)

A-W (Ahmed and Ward, 2002)

KMM (Saeed et al., 2001)

K3M

Table 3. Comparison of results: K3M and others.

Zhang and KMM
Guo and Ahmed and
Algorithm Wang (Saeed K3M
Hall (1992) Ward (2002)
(1996) et al., 2001)
1-pixel wide – + – + +
bias-reduced skeleton (preserves right angles) – – – – +
preserves topology (no disappearances) – + – + +

1. The algorithm preserves right angles at the lines in- Despite the fact that these drawbacks do not affect
terconnections, which result in better correspondence the shape preservation, the authors are currently working
between the original and the modified image. on reducing the algorithm complexity and computational
power requirements. The new results will be the matter of
2. It produces a one-pixel-wide skeleton.
a future publication after proving their validity.
3. The iterative phases are clearly aimed at deleting pi-
xels in a specific neighbourhood, therefore a general
5. Parallelizing sequential algorithms
idea of this thinning method is easily found.
Non-one-pass iterative thinning algorithms (sequential al-
4. The universality of the algorithm allows wide and di-
gorithms with one or more subcycles) are not intrinsical-
versified applications.
ly suitable for parallel processing. Therefore, they tend to
Although the K3M algorithm has successfully pro- have difficulties with successful utilisation of modern pa-
ved its universal performance over many other algorithms, rallel data processing techniques like multi-core proces-
it is intrinsically iterative in nature, and hence it requires sors or processor grids (hardware or network implemen-
more computing power than other, non-iterative algori- ted). With the growing of the usage and popularity of pa-
thms. Moreover, each iteration contains seven sequential rallel processing hardware and its techniques, it is basic
phases. This increases the computational complexity and to develop solutions that would allow such algorithms to
makes the parallelization attempts difficult. profit from the use of multi-processors.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 330 K. Saeed et al.

In Fig. 24 one can see scanline-type sequential pro- dependence is non-existent or irrelevant. This is not
cessing, where rows are processed sequentially. For each always possible.
pixel, a decision is made based on the state of the neigh-
bours. • Division with overlapping regions—regular division
while providing indispensable information (informa-
tion dependence areas from the area of division) for
each parallel processor.

5.1. Division with overlapping regions. In the divi-

sion process, the image is divided into some overlapping
parts to be processed in parallel. These parts are called
regions. Figure 26 shows how an image space is divided
into two regions P1 and P2. Figure 27, however, shows
how they are overlapped.

Fig. 24. Sequential scanline-type processing.

P1
The steps of the iterative thinning algorithm are re-
peated several times in their consecutive iterations. Each
processing step may contain several sub-cycles with ope-
rations such as flagging some pixels (for example, as
candidates for deletion or as non-deletable pixels). Fur-
thermore, sub-cycles perform their operations depending, P2
among other things, on that information. The final deci-
sion depends on the flags of neighbours set in the previous
subcycles, making the algorithm impossible to perform in
a parallel way as there exists information dependence be-
tween neighbouring pixels (information on the actual state
of neighbours is required at the specified subcycle). Fig. 26. Parallel scanline-type processing on different image re-
The information dependence area grows with each gions P1 and P2.
subcycle included in the iteration, as the state of neighbo-
urs depends on the states of their neighbours, and so on. Obviously, in the region edge each pixel has its de-
The growth of the effective neighbourhood is presented in finite number of neighbours. We can then deduce that the
Fig. 25. neighbouring pixel weights have a basic role in the de-
cision taking process during the thinning procedure. The
edge pixels of P1 in the crossing border between P1 and
P2 actually stick the neighbours of other pixels belonging
to P2. That is why the edge pixels from two regions sha-
re neighbours and hence some useful data are available to
both parts.


Fig. 25. Dependence area growth due to information dependen-
ce.

There may be several solutions that allow parallel 

processing of such problematic thinning algorithms: 
• Modification of the thinning algorithm to diminish or
eliminate subcycles. In the case of elimination, it le-
ads to the possibility of fully parallel processing. In
the case of decreasing the number of subcycles, it di-
Fig. 27. Pixel in the region edge.
minishes the information dependence area.
• Division of the image in such a way that information In most algorithms (including KMM and K3M) the

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 331

neighborhood is of a 3 × 3 character, but the approach is 6. Conclusions and future work

also valid for 3×4 or 5×5 neighborhood basis algorithms.
The overlapping area serves as the container of infor- The first aspect presented in the paper was the survey pre-
mation needed for neighbourhood checking. The width of sentation of the state of the art in the thinning area. A very
the region depends on the number of subcycles in the thin- large number of algorithms in this field have been propo-
ning algorithm and the requested number of independent sed and due to the complicated nature of skeletonization
parallel cycles of the algorithm. Such a configuration may it is often unclear how the approaches are related to each
be presented as read/write privileges of processes (as seen other in terms of processing quality and the general idea.
in Fig. 28) assuming that the processes operate on a shared A number of important works ranging from the earliest
memory. to the latest publications on thinning were discussed and
compared. Short summaries with the available results of
those algorithms were also presented. For the sake of ge-
neral comparison, essential and basic traits of the algori-
P1 thms were shown. The main points of view were the ty-
P1 read, write pe of methodology used and the parallelization possibility
P2 read only (intrinsic or secondary) with final general remarks on the
P2 P1 read only obtained results or the dedicated application of the studied
algorithm. The survey is considered a background for the
P2 read, write
authors’ proposal of the thinning algorithm named K3M.
The second important aspect of the paper is the pre-
sentation of the authors’ algorithm evolved from the ori-
Fig. 28. Overlapping regions. ginal KMM algorithm. The proposed algorithm introdu-
ces its advantages in terms of several important characte-
Pseudocode of the non-parallelized thinning algorithm ristics: thinning quality, a large range of possible applica-
tions and the clarity of processing stages. The examples
while not thinned do shown in Section 4 for isolated letters, words and graphi-
foreach pixel in image cal symbols demonstrate the processing quality of K3M.
check pixel neighbours The examined examples showed that the K3M algorithm
do thinning routines presents constant quality for printed words, handwritten
end foreach
end do scripts, numbers and characters of both Latin and non-
Latin alphabets. This fact proves the universal character
Pseudocode of the parallelized thinning algorithm of the K3M algorithm compared with other algorithms
known to the authors, which can be considered the most
iteration := 1 important advantage.
while not thinned do
foreach pixel in image The next task, which has been proved to be true in
if neighbour in border band then practical applications, is that the idea of the algorithm is
while neighbour.iteration < iteration-1 do clear and simple to follow. This was mainly demonstrated
wait for synchronization
in the detailed description in Section 3. The clarity and
end while
end if simple nature of the algorithm make it easy to study and
check pixel neighbours use, which is significant when compared with other, often
do thinning routines empirically obtained, algorithms.
if pixel in border band then
send synchronization info The last important aspect shown and discussed in the
end if paper regards parallelization issues for sequential thinning
end foreach algorithms. This matter was considered in Section 5 and
iteration := iteration + 1
could certainly benefit in improving many sequential ite-
end do
rative thinning algorithms. It is particularly useful in the
context of the recent popularity of parallel processing har-
5.2. Prediction of the neighbouring processor deci- dware and software techniques.
sion. In some cases it is possible to predict the decision Future work is mainly aimed at two directions: the
that a neighbouring processor should make. In particular, first is related to fine-tuning of the algorithm in order to
if a pixel in the border area is “white,” no change can be eliminate the slight deformation of some lines currently
made in conventional thinning (with a possible exception occurring. The introduced changes, however, should not
of thinning algorithms that perform thinned shapes smo- hassle the universality of the applications of the algori-
othing). That may particularly simplify the synchroniza- thm. The second development direction is the reduction of
tion task as presented above. the number of phases in the iterative part of the K3M al-

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 332 K. Saeed et al.

gorithm. That would result in lowering the computational Chen, Y.-S. and Hsu, W.-H. (1988). A modified fast parallel
effort as well as improving the parallelization possibility algorithm for thinning digital patterns, Pattern Recognition
of the algorithm. Letters 7(2): 99–106.
Chen, Y.-S. and Hsu, W.-H. (1989a). A 1-subcycle parallel thin-
Acknowledgment ning algorithm for producing perfect 8-curves and obta-
ining isotropic skeleton of an l-shape pattern, Proceedings
This work is supported by the Rector of Białystok Tech- of the International Conference on Computer Vision and
nical University (grant no. W/WI/10/07). Pattern Recognition, San Diego, CA, USA, pp. 208–215.
The authors are indebted to the anonymous referees Chen, Y.-S. and Hsu, W.-H. (1989b). A systematic approach for
for their constructive and valuable comments, and for their designing 2-subcycle and pseudo 1-subcycle parallel thin-
extremely helpful suggestions that improved the quality of ning algorithms, Pattern Recognition 22(3): 267–282.
the presented work. Chen, Y.-S. and Hsu, W.-H. (1990). A comparison of some
one-pass parallel thinnings, Pattern Recognition Letters
References 11(1): 35–41.

Ahmed, M. and Ward, R. (2002). A rotation invariant rule- Chin, R.T., Wan, H.-K., Stover, D.L. and Iverson, R. D. (1987).
based thinning algorithm for character recognition, IEEE A one-pass thinning algorithm and its parallel implemen-
Transactions on Pattern Analysis and Machine Intelligen- tation, Computer Vision, Graphics, and Image Processing
ce 24(12): 1672–1678. 40(1): 30–40.

Altuwaijri, M.M. and Bayoumi, M.A. (1998). A thinning algori- Dinnen, G. (1955). Programming pattern recognition, Proce-
thm for arabic characters using art2 neural network, IEEE edings of the Western Joint Computer Conference, Los An-
Transactions on Circuits and Systems II: Analog and Digi- geles, CA, USA, pp. 94–100.
tal Signal Processing 45(2): 260–264. Dyer, C. and Rosenfeld, A. (1979). Thinning algorithms for
Ammann, C.J. and Sartori-Angus, A.G. (1985). Fast thinning gray-scale pictures, IEEE Transactions on Pattern Analy-
algorithm for binary images, Image Vision and Computing sis and Machine Intelligence 1(1): 88–89.
3(2): 71–79. Favre, A. and Keller, H. (1983). Parallel syntactic thinning by
Andreadis, I., Vardavoulia, M., Louverdis, G. and Papamarkos, recoding of binary pictures, Computer Vision, Graphics,
N. (2000). Colour image skeletonisation, Proceedings of and Image Processing 23(1): 99–112.
the 10th European Signal Processing Conference, Tampe- Gonzalez, R.C. and Woods, R.E. (2007). Digital Image Proces-
re, Finland, Vol. 4, pp. 2389–2392. sing, 3rd Edition, Prentice Hall, Upper Saddle River, NJ.
Arcelli, C. (1981). Pattern thinning by contour tracing, Compu- Guo, Z. and Hall, R.W. (1989). Parallel thinning with two-
ter Graphics Image Processing 17(2): 130–144. subiteration algorithms, Communications of the ACM
Arcelli, C. and di Baja, G.S. (1978). On the sequential appro- 32(3): 359–373.
ach to medial line transformation, IEEE Transactions on Guo, Z. and Hall, R.W. (1992). Fast fully parallel thinning algo-
Systems, Man and Cybernetics 8(2): 139–144. rithms, CVGIP: Image Understanding 55(3): 317–328.
Arcelli, C. and di Baja, G.S. (1981). A thinning algorithm ba- Hilditch, C.J. (1968). An application of graph theory in pattern
sed on prominence detection, Pattern Recognition 13(3): recognition, Machine Intelligence 3: 325–347.
225–235. Hilditch, C.J. (1969). Linear skeletons from square cupboards,
Arcelli, C. and di Baja, G.S. (1987). A contour characterization Machine Intelligence 4: 403–420.
for multiply connected figures, Pattern Recognition Letters Huang, L., Wan, G. and Liu, C. (2003). An improved paral-
6(4): 245–249. lel thinning algorithm, ICDAR ’03: Proceedings of the Se-
Arcelli, C. and di Baja, G.S. (1989). A one-pass two-operation venth International Conference on Document Analysis and
process to detect the skeletal pixels on the 4-distance trans- Recognition, Washington, DC, USA, pp. 780–783.
form, IEEE Transactions on Pattern Analysis and Machine Jaisimha, M.Y., Haralick, R.M. and Dori, D. (1994). Quantita-
Intelligence 11(4): 411–414. tive performance evaluation of thinning algorithms under
Baruch, O. (1988). Line thinning by line following, Pattern Re- noisy conditions, Proceedings of the IEEE Conference on
cognition Letters 8(4): 271–276. Computer Vision and Pattern Recognition (CVPR), Seattle,
Blum, H. (1967). A transformation for extracting new descrip- WA, USA, pp. 678–683.
tors of shape, in W.W. Dunn (Ed.), Models for the Percep- Jang, B.K. and Chin, R.T. (1992). One-pass parallel thinning:
tion of Speech and Visual Form, MIT Press, Cambridge, Analysis, properties and quantitative evaluation, Trans-
MA, pp. 362–380. actions on Pattern Analysis and Machine Intelligence
Chen, Y.-S. (1996). The use of hidden deletable pixel detec- 14(11): 1129–1140.
tion to obtain bias-reduced skeletons in parallel thinning, Ji, X. and Feng, J. (2004). A new approach to thinning ba-
ICPR ’96: Proceedings of the 13th International Con- sed on time-reversed heat conduction model, International
ference on Pattern Recognition, Washington, DC, USA, Conference on Image Processing (ICIP), Singapore, Vol. 1,
pp. 91–95. pp. 653–656.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 333

Ju, T., Baker, M.L. and Chiu, W. (2007). Computing a family Rockett, P. I. (2005). An improved rotation-invariant thinning
of skeletons of volumetric models for shape description, algorithm, IEEE Transactions on Pattern Analysis and Ma-
Computer-Aided Design 39(5): 352–360. chine Intelligence 27(10): 1671–1674.
Kirsch, R.A., Cahn, L., Ray, C. and Urban, G.H. (1958). Expe- Rosenfeld, A. (1975). A characterization of parallel thinning
riments in processing pictorial information with a digital algorithms, Information and Control 29(3): 286–291,
computer, IRE-ACM-AIEE ’57 (Eastern): Papers and Di- http://theory.lcs.mit.edu/
scussions Presented at the December 9-13, 1957, Eastern ~iandc/ic75.html
Joint Computer Conference: Computers with Deadlines to Rutovitz, D. (1966). Pattern recognition, Journal of Royal Stati-
Meet, New York, NY, USA, pp. 221–229. stical Society 129(4): 504–530.
Klette, R. and Rosenfeld, A. (2004). Digital Geometry: Geo- Saeed, K. (2001). Text and image processing: Non-interrupted
metric Methods for Digital Image Analysis, Morgan Kauf- skeletonization, Proceedings of the 1st International
mann, San Francisco, CA. IEEE Conference on Circuits, Systems, Comunications and
Lam, L., Lee, S.-W. and Suen, C.Y. (1992). Thinning Computers—IEEE-CSCC’01, Crete, Greece, Advances in
methodologies—A comprehensive survey, IEEE Trans- Scientific Computing, Computational Intelligence and Ap-
actions on Pattern Analysis and Machine Intelligence plications, World Scientific Engineering Society Press,
14(9): 869–885. pp. 350–354.
Lee, S.-W., Lam, L. and Suen, C. Y. (1991). Performance evalu- Saeed, K. (2004). Image Analysis for Object Recognition, Bia-
ation of skeletonizing algorithms for document image pro- łystok Technical University Press, Białystok.
cessing, Proceedings of the First International Conference Saeed, K. and Niedzielski, R. (1999). Experiments on thinning
on Document Analysis and Recognition, Saint-Malo, Fran- of cursive-style alphabets, International Conference on In-
ce, pp. 260–271. formation Technologies for Education, Science and Busi-
Malina, W., Ablemeyko, S. and Pawlak, W. (2002). Funda- ness ITESB’99, Minsk, Belarus, pp. 45–49.
mentals of Digital Image Processing, Akademicka Oficyna Saeed, K., Rybnik, M. and Tabedzki, M. (2001). Implementation
Wydawnicza EXIT, Warsaw, (in Polish). and advanced results on the non-interrupted skeletoniza-
Mallat, S. (1989). A theory for multiresolution signal decompo- tion algorithm, in W. Skarbek (Ed.) Computer Analysis of
sition: The wavelet representation, IEEE Transactions on Images and Patterns, Lecture Notes in Computer Science,
Pattern Analysis and Machine Intelligence 11(7): 674–693. Vol. 2124, Springer-Verlag, Heidelberg, pp. 601–609.
Parker, J., Jennings, C. and Molaro, D. (1994). A force-based Wan, Y., Yao, L., Xu, B. and Zeng, P. (2008). A distance map
thinning strategy with sub-pixel precision, Vision Interface based skeletonization algorithm and its application in fiber
Conference, Banff, Alberta, Canada. recognition, International Conference on Audio, Language
Pavlidis, T. (1980). A thinning algorithm for discrete binary and Image Processing, Shanghai, China, pp. 1769–1774.
images, Computer Graphics Image Processing 13(2): 142– You, X. and Tang, Y. Y. (2007). Wavelet-based approach to cha-
157. racter skeleton, IEEE Transactions on Image Processing
Pavlidis, T. (1981). A flexible parallel thinning algorithm, Proce- 16(5): 1220–1231.
edings of the International Conference on Pattern Recogni- Zhang, Y. Y. and Wang, P. P. (1996). A parallel thinning algo-
tion and Image Processing, Dallas, TX, USA, pp. 162–167. rithm with two-subiteration that generates one-pixel-wide
Pavlidis, T. (1982a). Algorithms for Graphics and Image Proces- skeletons, International Conference on Pattern Recogni-
sing, Springer, Computer Science Press, Rockville, MD. tion, Vienna, Austria, Vol. 4, pp. 457–461.

Pavlidis, T. (1982b). An asynchronous thinning algorithm, Com-

puter Graphics Image Processing 20(2): 133–157. Khalid Saeed received the B.Sc. degree in elec-
trical and electronic engineering in 1976 from
Pervouchine, V. and Leedham, G. (2005). Document exami- Baghdad University, the M.Sc. (electronics) and
ner feature extraction: Thinned vs. skeletonised handwri- Ph.D. (telecommunications) degrees from the
ting images, Proceedings of the IEEE Region 10 Technical Wrocław University of Technology, Poland, in
Conference (TENCON05), Melbourne, Australia, pp. 1–6. 1978 and 1981, respectively. He received his
D.Sc. degree (habilitation) in computer science
Quadros, W. R., Shimada, K. and Owen, S. J. (2004). 3d discrete from the Polish Academy of Sciences in Warsaw
skeleton generation by wave propagation on pr-octree for in 2007. From 1992 to 2008 he was with Biały-
finite element mesh sizing, SM ’04: Proceedings of the stok Technical University. Since 2008 he has be-
Ninth ACM Symposium on Solid Modeling and Applica- en a professor of computer science at the AGH University of Science and
tions, Aire-la-Ville, Switzerland, pp. 327–332. Technology in Poland. He has published more than 120 works, about 20
edited books, and seven text and reference books. His areas of interest
are image analysis and processing, biometrics and computer information
systems.

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

 334 K. Saeed et al.

Marek Tab˛edzki received his M.Sc. degree in param N - number of sticking neighbours needed
computer science from Białystok Technical Uni-
versity (Poland) in 2001. He is presently with {classic 1-stage}
the same university, at the Faculty of Computer
Science. His research interests include informa- mark image pixels as 1
tion processing systems, particularly digital ima- mark background pixels as 0
ge processing and pattern recognition. repeat
thinned = true
for each pixel in image
if pixel = 1 then
if (pixel.N mod 2 = 0) or
(pixel.S mod 2 = 0) or
Mariusz Rybnik received his M.Sc. degree in (pixel.W mod 2 = 0) or
computer science from Białystok Technical Uni- (pixel.E mod 2 = 0) then
versity (Poland) in 2001, and the Ph.D. degree pixel := 3
in engineering in biology and medicine from the end if
University of Paris XII—Val de Marne (France) end if
in 2004. He is currently an assistant professor at end for
the Faculty of Mathematics and Computer Scien- for each pixel in image
ce of the University of Białystok. His research if pixel = 3 then
interests include image processing, pattern reco- { DLA - deletion lookup array }
gnition, biometrics and artificial intelligence. if weight(pixel) in DLA then
pixel := 0
thinned = false
Marcin Adamski received his M.Sc. degree else
in computer science from Białystok Technical pixel := 3
University (Poland) in 2001, where he curren- end if
tly works at the Faculty of Computer Science. end if
His research interests concern image processing end for
and image recognition methods, and their appli- for each pixel in image
cations in biometric systems. if pixel = 3 then
pixel := 1
end if
end for
until thinned
Appendix mark even pixels as background
mark odd pixels as image
Pseudocodes of the implemented computer programs
{neighbour weight calculation}
{creating N-lookup arrays}
function weight(pixel) : integer
{alternate to weight calculation} return 1*(pixel.N mod 2) + 2*(pixel.NE mod 2)+
function lookup(N:integer) : boolean 64*(pixel.W mod 2) + 4*(pixel.E mod 2)+
vector = [pixel.N, pixel.NE, pixel.E, 32*(pixel.SW mod 2) + 16*(pixel.S mod 2)+
pixel.SE, pixel.S, pixel.SW, 8*(pixel.SE mod 2) + 128*(pixel.NW mod 2)
pixel.W, pixel.NW] end function
changes := 0
for i := 1 to 7 {single-for 1-stage}
if vector[i] <> vector[i+1] then
changes := changes + 1 mark image pixels as 1
end if mark background pixels as 0
end for phase := 1
if changes <= 2 then repeat
ones := 1 thinned = true
for i := 1 to 8 for each pixel p in image
if vector[i] = 1 then if p mod 2 = 1 then
ones := ones + 1 { checking if border pixel }
end if if (p.N <> 2*phase and p.N mod 2 = 0) or
end for (p.S <> 2*phase and p.S mod 2 = 0) or
if ones = N then (p.W <> 2*phase and p.W mod 2 = 0) or
return true (p.E <> 2*phase and p.E mod 2 = 0) then
end if { DLA - deletion lookup array }
end if if weight(p) in DLA then
return false p := 2 * phase
end function thinned = false
else
pixel.N - north neighbour of actual pixel p := 2 * phase + 1
pixel.NE - northeast neighbour of actual pixel end if
and so on... end if
end if

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC

K3M: A universal algorithm for image skeletonization and a review of thinning techniques 335

end for
phase := phase + 1
until thinned
mark even pixels as background
mark odd pixels as image

Received: 14 June 2009

Revised: 18 October 2009

Unauthentifiziert | Heruntergeladen 23.11.19 13:29 UTC