AN IMPROVED APPROACH FOR THINNING BY PRESERVING

LOCAL COUPLING POINTS

Lalit Kumar, Abhay Bansal, Dinesh Ganotra and Neeraj Jain

Newgen Software Technologies Limited, A-6, Satsang Vihar Marg, Qutab Institutional Area, New Delhi–110067 India

Keywords: Binarization, Thinning, Skeleton, Local Coupling Point, LCP, ICR, OCR, Offline Signature Verification.

Abstract: Image thinning, while preserving geometrical properties of the image, remains one of the most challenging

areas in Image Processing. In this paper, we introduce an image thinning approach that takes as input gray

scale images and produces binary thinned images as output. Our approach uses bi-directional iterative

thinning process to get single-pixel-thinned image (primary skeleton), while preserving the geometric

properties of the image. The gray scale image is binarized using dyamic threshold, and further processed to

remove noise. Black pixels are classified as contour pixels (pixels exposed to the background) and body

pixels (black pixels other than the contour pixels). Thinning process involves scanning contour pixels from

two opposite directions simultaneously, while preserving Local Coupling Points (LCP) and removing the

rest of pixels.

1 INTRODUCTION

Thinning of images is a fundamental requirement in

a wide range of application areas, such as Courtesy

Amount Recognition and Legal Amount

Recognition (CAR, LAR) (

Lecce et al., 2000),

Handwritten Signature verification (Ammar et al.,

1986), Optical Character Recognition(Cowell and

Hussain, 2003), Fingerprint Classification (Mehtre,

1993), Printed Circuit Boards (El Mesbahi and

Chaibi, 1993), etc.

A number of algorithms implementing various

approaches for thinning of images have been

proposed. The algorithms are broadly divided into

two categories: iterative and non-iterative. Iterative

algorithms repeatedly apply a procedure on the input

image till the skeleton image is obtained. On the

contrary, non-iterative algorithms take a set of

predetermined parameters and try to obtain primary

skeleton in a single iteration (Neuslds and

Olszewski, 1994).

The iterative algorithms are further divided into

two sub-categories, sequential and parallel. In

sequential algorithms all the pixels are considered to

take decision for a single pixel. On the other hand, in

parallel algorithms (Ubeda, 1993), (Hang and Wang,

1994) individual pixels can be independently judged

for decision-making. The proposed algorithm is a

combination of sequential and parallel approach.

In the past decade, a number of algorithms have

been proposed for thinning of gray scale images.

Most of these deal with simple images, such as

handwritten characters and machine-printed text in

specific languages (Ahmed et al., 2002). Few

approaches effectively deal with complex images

such as signatures while at the same time preserving

the intricate shapes of such images. For example,

both Han’s algorithm (Datta and Parui, 1994) and

Datta’s algorithm (Han et al., 1997) are able to

preserve the shapes of the simple objects, but split

the edges into spurious branches. Shape preservation

is a key requirement for off-line signature

verification. This paper targets the thinning of

signature image while preserving geometrical

features, such as long strokes, curves cross points,

and end points. With threshold changes, the

proposed approach would find extensive use in

application areas, such as detecting the CAR-LAR

amount and enhancing the ICR engines’ accuracy in

reading text written with thick pen.

The paper is organized as follows: Section 2.1

deals with preprocessing that includes binarization

of image and noise removal. Section 2.2 deals with

the classification of the black pixels as Local

Coupling Point (LCP) and body pixels. In section 3,

experimental results are discussed.

Kumar L., Bansal A., Ganotra D. and Jain N.

AN IMPROVED APPROACH FOR THINNING BY PRESERVING LOCAL COUPLING POINTS.

DOI: 10.5220/0002156700900095

In Proceedings of the Fourth International Conference on Computer Vision Theory and Applications (VISIGRAPP 2009), page

ISBN: 978-989-8111-69-2

The following conventions are used in the paper:

G Graph

M Matrix

n Number of elements

V One dimensional vector

(x,y)

Element at xth column and yth row of

matrix M

2 PROPOSED ALGORITHM

The proposed algorithm uses an iterative thinning

process that preserves the geometrical features in

complex images. The algorithm is based on the

concept of classifying pixels as LCP and preserving

them.

For understanding LCP, consider a graph G

the form of an nxn matrix, where nodes are

represented by black pixels and edges by connected

black pixels. For a 3×3 matrix, there exists only one

interior node, M

(2,2),

connected to all exterior nodes.

We define an LCP as an interior node, which

when removed causes discontinuity in graph, G; i.e.,

a path, which traces all the exterior nodes, cannot be

established without including the LCP.

Figure 1(a)(i) represent the 3x3 pixels matrix, M,

where M

(2,2)

is the pixel to be checked for LCP;

(a)(ii) represents M in the form of a graph where

(1,1)

, M

(2,2)

, M

(3,1)

, and M

(3,3)

corresponds to 1, c, 2,

and 3 nodes of graph G

respectively. If node c is

removed, as shown in (a)(iii), there is no connection

between nodes 1, 2, and 3. So, M

(2,2)

is an LCP. On

the other hand, in Figure 1(b)(iii), even if we remove

node c, there exists a path connecting nodes 1, 2, 3

and 4, and therefore M

(2,2)

is not an LCP.

(i) (ii) (iii)

(a)

(i) (ii) (iii)

(b)

Figure 1: (i) 3×3 Pixel matrices, M; (ii) Graphs of matrices

represented as nodes and edges; (iii) Graphs of matrices

with nodes ‘c’ removed.

2.1 Preprocessing

All gray scale images are binarized with the use of

the modified Niblack algorithm (Chandra et al.,

2007). This algorithm works fine for handwritten-

text images. Once we obtain the binarized image, a

noise removal algorithm is applied to clean the

spatial noise (small components having size of 1-3

pixels). A simple 5x5 morphological filter is applied

to achieve this.

2.2 Identification of Local Coupling

Points

Identifying an LCP involves the following steps:

i. Calculate the number of adjacent Black&-

White-pixels pairs.

ii. If 3×3 pixels matrix has more than one

adjacent black-&-white-pixels pair, then

apply template matching. Template

matching is explained in section 2.2.2.

2.2.1 Black-&-White-Pixel Pairs

In a 3×3 mask, we represent all the elements of

(x,y),

except M

(2,2),

as a vector, V

, as shown in

Figure 2. Next, we count the number of transitions

from ‘0’ to ‘1’ or ‘1’ to ‘0’ (‘1’ represents black and

‘0’ represents white) considering each pixel only

once. As shown in figure 2, there are three

transitions.

(i)

0 1 1 0 1 0 1 1

(ii)

Figure 2: (i) 3x3 Pixels matrix, M; (ii) Vector

representation, V

of M.

2.2.2 Template Matching

Template matching for M

(x,y)

can be done by two

methods. In the first method, an exhaustive search is

done, where M

(x,y)

is matched

with predefined

templates, pixel by pixel, as shown in Figure 3. If a

perfect match is found, M

(2,2)

is not an LCP. Such an

exhaustive search is not efficient. Improvement can

be achieved by categorizing the templates on the

basis of number of black pixels that surround M

(2,2).

2 3

3 4

AN IMPROVED APPROACH FOR THINNING BY PRESERVING LOCAL COUPLING POINTS

Figure 3: LCP templates based on the number of

connected pixels; first set of templates with two black

pixels connected to the central pixel, second set with three

connected pixel, and so on.

The cases where seven surrounding pixels are

connected to the central pixel is not considered

because for such a case all the possible templates

will not be LCP.

The alternate method is to represent M(x,y), in

the form of a number between 0-255, and check the

value in the Lookup table Li (Table 1) to determine

whether the pixel under consideration is an LCP.

This is discussed next.

Assume that M(x,y)=0 denotes a white pixel, and

M(x,y)=1 denotes a black pixel. Represent exterior

pixels of M(x,y), using byte data type, B, where

each bit, bi, of B corresponds to the exterior pixel of

M(x,y), that is b1, b2, b3, b4, b5, b6, b7, and b8

correspond to M(1,1), M(1,2), M(1,3), M(2,3),

M(3,3), M(3,2), M(3,1), and M(2,1) respectively.

Calculate the decimal equivalent, D, of B. Look for

the value at Dth index in Lookup Table, Li. If LD is

1, M(2,2) is an LCP, otherwise not. For example, as

shown in Figure 2, binary representation of B is

01101011

, and equivalent decimal representation,

D, is 10710

. As L107 is 0, M(2,2) is not an LCP.

For thinning, all the pixels, which lie on the

contour of the image, are removed. For the strokes

that have single pixel width, single-pixel lines are

preserved to prevent any discontinuity. This is

discussed next.

In the first step, all the black pixels are classified

either as boundary pixels or body pixels.

i. A Boundary pixel is one that is exposed to the

background, i.e., a pixel having at least one white

pixel among eight neighbourhood pixels (Top,

Bottom, Left, Right, Top-Right, Right-Bottom,

Bottom-Left, and Left-Right).

ii. A Body pixel is one that is completely

surrounded by the eight black pixels in its immediate

neighbourhood.

In the second step, LCPs are identified among

boundary pixels. All Boundary pixels and LCPs are

labelled as black pixels, as shown in Figure 5 and

Figure 6 respectively.

The third step involves performing a row-wise

scan, and removing those boundary pixels, which are

adjacent to the body pixel and are not an LCP.

These three steps are performed iteratively until

no body pixels can be found, or further boundary

pixels cannot be removed.

The pseudo code for this algorithm is presented

below:

Algorithm1: Basic thinning pseudo code.

Mark all the boundary pixels and body pixels

While ((body pixels exist) And (Boundary pixels can be

removed))

If pixel under consider (is adjacent to body pixel)

And (Not a LCP) Then

Remove the pixel under consideration.

Else

Do not remove the pixel

End If

End While

(a)

(b)

Figure 4: (a) Input image (b) Thinned image.

The method results in a thinned image as shown

in Figure 4. However, some applications need one-

pixel-thick image that can be segmented in disjoint

polylines. This is discussed next.

For a pixel to be categorised as body pixel, the

constraints are relaxed. If a pixel is surrounded by

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

three black pixels from four immediate

neighbourhood pixels (Top, Bottom, Left, Right), it

is categorised as a body pixel. Subsequently, T-

shaped body pixels are obtained, as shown in Figure

5(a).

Next, an iterative row scan is performed, and

boundary pixels, adjacent to T-shaped body pixels,

are removed until no T-shaped body pixel exists or

further boundary pixels cannot be removed.

The more the body pixel condition is relaxed,

more thinned image is obtained. If the condition is

relaxed so that a pixel is considered a body pixel if it

is connected to two neighbourhood pixels (one pixel

at Top or Bottom, and other at Left or Right), L-

shaped pixels are obtained, as shown in Figure 5(b).

In this manner, a single-pixel-thinned image is

obtained.

(a)

(b)

Figure 5: (a) T-shaped pixels (b) L-shaped pixels.

Algorithm 2: Single-pixel thinning pseudo code.

Mark all the boundary pixels and body pixels

b_pixels Å body pixels

While ((b_pixels exist) And (Boundary pixels can be

removed))

If pixel under consideration (is adjacent to body pixel)

And (Not an LCP) Then

Remove the pixel under consideration.

Else

Do not remove the pixel

End If

End While

Mark all the boundary pixels and T-shaped pixels

b_pixels Å T-shaped pixels

Go to step2

If (no more boundary pixels can be removed)

Remove all the T-shaped body pixels

End If

Mark all the boundary pixels and L-shaped pixels

b_pixels Å L-shaped pixels

Go to step2

If (no more boundary pixels can be removed)

Remove all the L-shaped body pixels

End If

(a)

(b) (c)

Figure 6: (a) Original image; (b) Boundary pixels, marked

as black, are checked whether they are LCP. If a boundary

pixel is not a LCP, then it is removed; (c) LCP pixels

marked as black, and rest of the boundary pixels to be

removed marked as gray. The complete process is

repeatedly performed until non-LCP boundary pixels

exist.

(i) (ii)

(iii) (iv)

(a)

(i) (ii)

(iii) (iv) (v)

(vi) (vii) (viii)

(ix) (b) (c)

Figure 7: (a) Pass I, steps (i) to (iv); (b) Pass II, steps (i) to

(ix); (c) Pass III.

AN IMPROVED APPROACH FOR THINNING BY PRESERVING LOCAL COUPLING POINTS

The results, for every step of each pass of the

algorithm on a signature image, are shown in Figure

7. As can be seen, the algorithm takes four iterations

to complete Pass I. Pass II consist of nine iterations

to remove T-shaped boundary pixels. Finally, pass

III takes a single iteration to remove L-shaped

pixels.

3 EXPERIMENTAL RESULTS

Thinning algorithms are usually evaluated on four

parameters: connectivity preservation, skeleton

width, robustness to noise, and information

preservation. In addition to the above parameters, we

consider some other important aspects, such as

preservation of edges and robustness of the

algorithm to complex structures.

A large number of thinning algorithms that have

been proposed are inflicted with basic problems. For

example, the output of Datta’s algorithm (Datta and

Parui, 1994) has spurious branches. Another

example is Han’s algorithm (Han et al., 1997),

which certainly is an improvement over Datta’s

algorithm as it can remove spurious branches, but

skeleton width is the major problem. Moreover the

test patterns taken for Han’s are simple machine

printed characters that are less prone to noise and

irregular shapes at the ends. Further, the algorithm

by Lei Huang (Hunag et al., 2003) effectively

handles both the above problems, however, it is not

able to preserve the shapes of more complex images,

such as handwritten signatures. It also does not offer

user the flexibility of selecting the skeleton width.

The approach presented in our paper solves a

number of potential problems including those

mentioned previously, such as spurious branches,

noise, shape preservation, etc., while at the same

time it effectively and efficiently addresses

requirements of custom skeleton width selection and

thinning of complex structures.

In Figure 8, some handwritten signature images

and their corresponding thinned images, obtained by

our method, are shown.

The proposed algorithm is highly efficient. It is

easy to implement using programming languages,

such as C, which support bit-wise operators. The

algorithm’s superior performance is due to the use of

two unique approaches, a) Vector representation

followed by matching with look-up table, and b)

template matching. Leveraging these two

techniques, the algorithm provides unmatched

flexibility in processing images containing

handwritten signatures. The method has been tested

Figure 8: Original and thinned images.

with Grupo de Procesado Digital de Senales”

(GPDS) Signature Database (2016 images)

(Martinez et al., 2004). The average time taken for

the random set of 100 images (300x250) on Pentium

IV 2.5 GHz, 512MB RAM is approximately 9

seconds.

4 CONCLUSIONS

This paper presents an approach for thinning as well

as shape preservation of more complex structures,

such as handwritten signatures. The result presented

here clearly demonstrates that the proposed

method/algorithm is not only time efficient but also

preserves information.

REFERENCES

Ahmed, S., Sharmin, M. & Chowdhury Mofizur Rahman,

2002. A Generic Thinning Algorithm with Better

Performance, Fifth International Conference on

Computer and Information Technology.

Ammar, M., Yoshida, Y. & Fukumura T., 1986. A new

effective approach for off-line verification of

signatures by using pressure features. In: Proceedings

of the International Conference on Pattern

recognition.

Chandra, Lal, Lal, Puja, Gupta, Raju, Tayal, Arun &

Ganotra, Dinesh, 2007. Improved Adaptive

Binarization Technique for Document Image Analysis,

2nd International Conference on Computer Vision

Theory and Applications.

Cowell, Dr John, Hussain, Dr. Fiaz, 2003, Amharic

Character Recognition using a Fast Signature Based

Algorithm, Proceedings of the Seventh International

Conference on Information Visualization.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

Datta, A. & Parui, S.K., 1994. A Robust Parallel Thinning

Algorithm for Binary Images. Pattern Recognition,

Vol. 27, No. 9 pp.1181-1192.

El Mesbahi, J., Chaibi, M, 1993. Printed circuit boards

inspection using two new algorithms ofdilatation and

connectivity preserving shrinking, Neural Networks

for Signal Processing (1993) III. Proceedings of the

1993 IEEE-SP Workshop.

Han, N.H., La, C.W. & Rhee, P.K., 1997. An Efficient

Fully Parallel Thinning Algorithm, Proceedings of the

4th International Conference on Document Analysis

and Recognition, Page 137, ISBN 0-8186-7898-4.

Hang, Z. & Wang, Y.Y., 1994. A New ParallelThinning

Methodology, IJPRAI(8), pp.999-1011.

Hunag, Lei, Huang, Genxun & Liu, Chnagping, 2003. An

improved parallel thinning algorithm. Proceedings 7th

ICDAR Conference.

Lecce, V.Di, Dimauro, G., Guerriero, A., Impedovo, S.,

Pirlo, G. & Salzo, A., 2000, A new hybrid approach

for legal amount recognition, In: L.R.B. Schomaker

and L.G. Vuurpijl (Eds.), Proceedings of the Seventh

International Workshop on Frontiers in Handwriting

Recognition, Amsterdam, ISBN 90-76942-01-3,

Nijmegen: International Unipen Foundation, pp 199-

208.

Martinez, L.E., Travieso, C.M., Alonso, J.B. & Ferrer, M.,

2004. Parametrization of a forgery Handwritten

Signature Verification using SVM, IEEE 38th Annual

2004 International Carnahan Conference on Security

Technology, pp. 193-196.

Mehtre, B., 1993. Fingerprint image analysis for automatic

identification, Machine Vision and Applications, vol.

6, pp. 124–139.

Neuslds, Christian & Olszewski, Jan, 1994. A noniterative

Thinning Algorithm, ACM Transactions on

Mathematical Software, Vol. 20, No. 1, March 1994,

Pages 5-20.

Ubeda, V.,1993. A Parallel Thinning Algorithm Using

KxK Masks, IJPRAI(7), pp. 1183-1202.

APPENDIX

Table 1: Detailed Lookup table snippet.

Table 2: Detailed Lookup table snippet.

Decimal Representation Is LCP

0-2, 4, 8-11, 16-19, 25-27, 32-47, 49-51,

57-59, 64, 66, 68, 70, 72-76, 78, 80, 82,

85, 88, 90, 96, 98, 100, 102, 104-108,

110, 114, 117, 121-122, 127-128, 130,

132, 134, 136-140, 142, 144-148, 150,

152-156, 158, 160-180, 182, 184-188,

190-191, 194, 196, 198, 200-204, 206,

210, 218, 223, 226, 228-232, 242, 250

3, 5-7, 12-15, 20-24, 28-31, 48, 52-56,

60-63, 65, 67, 69, 71, 77, 79, 81, 83-84,

86-87, 89, 91-95, 97, 99, 101, 103, 109,

111-113, 115-116, 118-120, 123-126,

129, 131, 133, 135, 141, 143, 149, 151,

157, 159, 181, 183, 189, 192-193, 195,

197, 199, 205, 207-209, 211-217, 219-

222, 224-225, 227, 233-241, 243-249,

251-255

√

AN IMPROVED APPROACH FOR THINNING BY PRESERVING LOCAL COUPLING POINTS