ISBN 978-952-5726-07-7 (Print), 978-952-5726-08-4 (CD-ROM)

Proceedings of the Second Symposium International Computer Science and Computational Technology(ISCSCT ’09)
Huangshan, P. R. China, 26-28,Dec. 2009, pp. 281-284

Analysis and Application of Iteration

Skeletonization Algorithm in Recognizing
Chinese Characters Image
Tingmei Wang1, Ge Chen2, and Zhansheng Chen2
1
Applied Science and Technology Institute Beijing Union University, Beijing, China
Email: [email protected]
2
Applied Science and Technology Institute Beijing Union University, Beijing, China
Email: [email protected], [email protected]

Abstract—The paper studied several image skeleton several classical iterative skeleton extraction algorithms in
extraction algorithms, such as Zhang-Suen, Rosenfeld and the character image skeleton extraction applications.
Pavlidis etc. And compared extraction effects based on
different Chinese characters by making computer programs, II. MODELS OF IMAGE SKELETON EXTRACTION
and gave advice how to choose the best algorithm for
recognizing a certain Chinese characters image, and then A. Fire Spreading
gave a conclusion that Zhang-Suen and Rosenfeld
algorithms were the best algorithms for extracting Chinese Fire spreading refers to all border points of image are
characters skeletons. lighted at the moment of t=0, the flame spreads towards
internal of image at the same speed. When the meet
Index Terms—Chinese Characters Image Processing; Image happens in front of wave, the flame goes out. Collection
Skeleton Extraction; Zhang-Suen; RosenFeld; Pavlidis of points where flame goes out constitutes axis named
image skeleton denoted by SKF shown as fig.1
I. INTRODUCTION B. Maximum Disc
With development of Internet and application of Maximum disc indicates that suppose D is one of
computer technology, a plenty of information exists in inscribed disc of image, this means at least two points of
form of images. Retrieving among so many images the disc are tangent to image edge. If D is not a subset of
becomes very difficult. When retrieving Image based on other inscribed discs within image, D is the greatest disc
key words, we must add mark for them so that efficiency as shown in fig.2. Now image skeleton denoted by SKM
of retrieval is very low. In this case, image retrieval based can be defined as collection of the greatest discs within
on content (CBIR) which use color, strip, shape, space image.
relation etc, is needed. Image retrieval based on space
relation uses space topology structure of objects in images
to retrieve. Image skeleton is one of the best effective
ways to represent the topology structure. It is widely used
in shape description, pattern recognition, and industrial
inspection, and image compression coding etc. Image
Figure 1. Fire spread skeleton. Figure 2. Maximum disc skeleton
skeleton was proposed by Blum [1]. He represented image
skeleton by means of axis. C. Peeling Method
Image skeleton extraction algorithm generally
Peeling method indicates that staring from edge of line,
possesses the following characteristics: The first is you peels it off one layer with one pixel width at a time
connectivity, which means connectivity of image skeleton till the single pixel lines that are connected each other
must be consistent with original image. The second is and constructed by single raster are obtained. One
thinning, which means width of image skeleton should be condition should be considered during peeling because a
one pixel. The third is axis, which means image skeleton line in different locations may have different widths, that
should be as far as possible the center line of original is to say the pixel that may result in lines disconnected
image. The fourth is maintainability, which means image can never be peeled. The basic principle of the algorithm
skeleton should retain as far as possible detail of original is that peeling pixels that have no effect on connectivity
image. Last one is rapidity, which means running speed of image raster topology, and conversely these pixels
of algorithm should be as fast as possible. should be retained.
Image skeleton extraction in Chinese character
recognition is an important research topic. Image skeleton III. ANALYSIS OF SEVERAL CLASSICAL IMAGE SKELETON
extraction algorithm consists of iterative algorithm and EXTRACTION ALGORITHMS BASED ON PEELING
non-iterative algorithm. This paper mainly discusses
© 2009 ACADEMY PUBLISHER
AP-PROC-CS-09CN005 281
A. Principle of Template Matching Algorithm If 8-neighborhood of the lower edge point of P
Eight points adjacent to pixel p are called 8- matches templates 1,3,4,5,6,7,8, continue scanning.
neighborhood of pixel P as shown in fig.3, and it can be If 8-neighborhood of the left edge point of P
denoted by Pi such that the value of i is less than 1 and matches templates 1,2,4,5,6,7,8, continue scanning.
greater than 8. P1,P3,P5 and P7 are called four-adjacent If 8-neighborhood of the right edge point of P
area point, and P1,P2,P3,P4,P5,P6,P7,P8 are called eight- matches templates 1,2,3,5,6,7,8, continue scanning.
adjacent area point. Otherwise mark point P for deleting.
Template matching algorithm uses a template 3*3 to }
retrieve images and judges if there are points being the }
same case as the template. If there are such points, you C. Zhang-Suen Algorithm
should execute corresponding operations according to the Zhang-Suen algorithm was advanced first by Zhang T
operations of template such as retaining operation, Y in 1984 [3]. It is also an image skeleton extraction
deletion etc. In short, Rosenfeld, Zhang-Suan and algorithm based on parallel template matching and
Pavlidis algorithms are image skeleton extraction deletion. The algorithm defines condition for points to be
algorithm based on template matching. deleted as shown in fig.5. For those points satisfying
deleting condition, you should delete them. As shown in
fig.5, if satisfying one black point or two gray points
representing background color, you delete the point.
Processing cycle of the algorithm consists of two steps.
Finding a point matching with the template shown in
Figure 3. 8- neighborhood of pixel P.
fig.5(1) is the first step, the second step is to verify the
point matching with the template shown in fig.5(2),then
B. Rosenfeld Algorithm repeating the above process until there are no points to be
Rosenfeld algorithm was first advanced by Stefanelli R
and Rosenfeld A in 1971[2].It is an image skeleton
extraction algorithm based on parallel template matching.
The algorithm defines condition for final pixel as shown
in fig.4 and judges if edge pixel satisfies the condition so
Figure 5. Zhang-Suen chinning template
as to determine how to operate it. As shown in fig.4,
there exists at least one black point among these points
deleted in a certain cycle.
marked X, and the same is true for Y. here, black point
Image skeleton extraction procedure of Zhang-Suen
represents foreground color, and white point represents
algorithm can be described as follows:
background color, and gray point represents foreground
while (there still exist points to be deleted)
or background colors. You should retain pixel points
{
satisfying final pixel condition, otherwise remove them.
For each foreground point P search:
Processing cycle of the algorithm consists of four main
{
steps: completing a removal of edge pixels from four
If 8-neighborhood of edge point of P matches
directions-moving up, down, left or right, then repeating
template 1, delete point P.
this process until no deleting occurs in a certain cycle.
If eight adjacent area of edge point of P matches
Now process of image skeleton extraction is finished.
template 2, delete point P.
}
}
D. Pavlisis Algorithm
Palidis algorithm was advanced first by Pavlidis in
1982[4]. It is an image skeleton extraction algorithm
based on template matching and retaining. The algorithm
defines the condition for points to be retained showed in
Figure 4. Final pixel condition
fig.6. For those points satisfying retaining condition, you
should retain them. Processing cycle of the algorithm
Image skeleton extraction procedure can be described consists of four steps. The first step is to judge if point of
as follows: right edge matches the template, then repeat the same
while(there still exist pixels to be deleted) operation to process its upper edge, left edge and lower
{ edge until there are no points to be deleted. Now process
For foreground color point P search: of image skeleton extraction is finished.
{ Image skeleton extraction of Pavlidis was described as
If 8-neighborhood of the upper edge point of P follows:
matches templates 2,3,4,5,6,7,8, continue scanning. while(there still exist points to be deleted)
{

282
 For each foreground point search: obtain image skeleton of single pixel connecting by
{ peeling edge pixels many times. These algorithms have
If 8-neighborhood of right edge point of P different times of iteration. One iteration cycle of Pavlidis
matches templates 1,2,3,4,5,6, continue scanning. algorithm and Rosenfeld algorithm consists of four steps,
If 8-neighborhood of the upper edge point of P and Zhang-Suen algorithm consists of two steps in one
matches templates 1,2,3,4,5,6, continue scanning. iteration cycle. At the same time, different iteration
If 8-neighborhood of the left edge point of P processes lead to different number of templates. Zhang-
matches templates 1,2,3,4,5,6, continue scanning. Suen algorithm has the least number of templates so that
If 8-neighborhood of lower edge point of P its speed is the fastest. The number of templates and the
matches templates 1,2,3,4,5,6, continue scanning. times of iterations have effect on efficiency of algorithms.
Otherwise mark the point P for deletion. The less the algorithm uses the number of templates and
} the less the algorithm has the times of iteration, the faster
} the algorithm runs.

IV. EXPERIMENT RESULT AND ANALYSIS

Programming and running these algorithms including
Hilditch, Pavlidis, Rosenfeld, Naccache, Deutsch and
Zhang-Suen in turns to extract image skeleton for
handwritten Chinese characters, printed Chinese
characters and Chinese calligraphy as shown in fig.7 and
Figure 6. The condition of retaining point(At lease one of X,Y is 1) processing results are shown in fig.8,fig.9 and fig.10.

E. Other Image Skeleton Extraction Algorithms

The deleting condition of Hilditch algorithm and
retaining condition of Pavlidis algorithm are
complementary [5]. Similar image skeleton using these
two algorithms can be obtained, so their description will
not be discussed in detail. Naccache algorithm marks
safety point including either breakpoint or endpoint to be
Figure 7. Experiment images
retained and otherwise these points are to be removed,
then deletes all the points described as removed until
there is no point to be deleted.
Deutsch algorithm is an image skeleton extraction
algorithm based on deletion. It defines some deletion
conditions, and if there exist such points that satisfy
deletion condition, the points should be deleted.
Otherwise the points should be retained. The algorithm
extracts image skeleton by two cycles executed
alternately.
F. Comparison of Skeleton Extraction Algorithms
From the discussion presented above, we can know
that these algorithms mentioned obtain image skeleton by Figure 8. Handwritten Chinese characters
conditions matching and many times iterations. Some of
them extract image skeleton through given conditions,
such as Hilditch algorithm and Deutsch algorithm. Some
of them extract image skeleton through templates
matching, such as Zhang-Suen algorithm and Rosenfeld
algorithm. Hilditch algorithm, Deutsch algorithm and
Zhang-Suen algorithm extract image skeleton based on
deletion, and Rosenfeld algorithm, Pavlidis algorithm and
Naccache algorithm extract image skeleton based on
retaining. Though reservation conditions are called by
different names in these algorithms, final pixel condition
in Rosenfeld, retaining point condition in Pavlidis, and
safety point condition in Naccache, but their principles
are the same. Figure 9. Printed Chinese characters
All the skeleton extraction algorithms mentioned
above are based on peeling. In other words, we can

283
 • When processing handwritten Chinese characters,
Deutsch algorithm can get good effect, but if
processing printed Chinese characters and
Chinese calligraphy, the effect of it is not
satisfactory, because there are not only many
superfluous branches, but also a little deformity.

V. CONCLUSIONS
The paper introduces some classical models of image
skeleton extraction, analyzes principles of several
classical image skeleton extraction algorithms based on
Figure 10. Chinese calligraphy
peeling and tests these algorithms through processing
Chinese calligraphy, handwritten Chinese characters and
Drawing a comparison for experimental figures, printed Chinese characters. Experiment result shows that
conclusion is shown as follows: efficiency of Zhang-Suen algorithm is higher than that of
• If you want to recognize Chinese calligraphy, other algorithms. Using Rosenfeld and Zhang-Suen
algorithms to process Chinese characters, we can obtain
handwritten Chinese characters and printed
image skeleton with a good topology features and
Chinese characters, the processing effect of
without superfluous branches. So a conclusion can be
Rosenfeld algorithm is better than Hilditch,
drawn that Zhang-Suen and Rosenfeld algorithms are the
Pavlidis, Daccache and Deutsch algorithms’.
best algorithms for extracting Chinese characters
• Both Hilditch algorithm and Pavlidis algorithm
skeletons.
can produce similar image skeletons. They can
result in good effect for processing handwritten
REFERENCES
Chinese characters and printed Chinese characters.
But when it comes to processing Chinese [1] Blum H. A transformation for extracting new description
calligraphy, they can not process the intersection of shape. In: Wathen-Dunn W ed. Model for the
of lines accurately. Perception of Speech and Visual. Cambride, Massachusetts:
MIT Press, 1967,pp 362-380.
• Considering image skeleton connectivity, Pavlidis
[2] Stefanelli R, Rosenfeld A.Some parallel thinning
algorithm is worse than Hilditch. When algorithms for digital pictures[J].Journal of ACM,1971(2).
processing Chinese calligraphy by Pavlidis [3] Zhang T Y. A fast parallel algorithm for thinning digital
algorithm, image skeleton exists apparent fracture patterns[J]. Communications of ACM, 1984, 27(3):236-
shown as fig.11. 239.
• Using Rosenfeld and Zhang-Suen algorithms, we [4] Pavlidis T.Algorithms for Graphics and lmage
can obtain image skeleton with no superfluous Procesing[M] . Rockville:Computer Science Press,1982.
branches, and they can process intersection of [5] Hilditch C J.Linear Skeletons from Square
lines accurately. Cupboards[J].Machine Intelligence, 1969;4:403.
[6] Naccache N J, Shinghal R. SPTA: A proposed algorithm
• When processing handwritten Chinese characters for thinning by binary patterns[J]. IEEE Transactions on
and printed English characters, the processing System, Man Cybernetics, 1984, 14(3):409-418.
effect of Naccache algorithm is better than that of [7] Deutsch E S. Thinning algorithms on rectangular
others’. But for processing printed Chinese hexagonal and triangular arrays[J].Communications of the
characters and Chinese calligraphy, the effect is ACM,1972,15(9):827～837.
not satisfactory, because there exist so many
superfluous branches.

284