Graph based Method for Online Handwritten Character Recognition

Rabiaa Zitouni

1 a

, Hala Bezine

and Najet Arous

Laboratory LR-SITI ENIT, University Tunis El Manar, B.P.37, 1002 Tunis, Tunisia

Laboratory REGIM ENIS, University Sfax, B.P.1173, 3038 Sfax, Tunisia

Keywords:

Fuzzy Attributed Relational Graph, Graph Matching, Structural Pattern Recognition, Handwritten Graphs,

Tree Search Method.

Abstract:

In this research, we attempt to propose a novel graph-based approach for online handwritten character recog-

nition. Unlike the most well-known online handwritten recognition methods, which are based on statistical

representations, we set forward a new approach based on structural representation to overcome the inherent

deformations of handwritten characters. An Attributed Relational Graph (ARG) is dedicated to allowing the

direct labeling of nodes (strokes) and edges (relationships) of a graph to model the input character. Each node

is characterized by a set of fuzzy membership degrees describing their properties (type, size). Fuzzy descrip-

tion is invested in order to guarantee more robustness against uncertainty, ambiguity and vagueness. ARGs

edges stand for spatial relationships between different strokes. At a subsequent stage, a tree-search based

optimal matching algorithm is explored, which allows the search for character structures i.e the minimum cost

of nodes. Experiments performed on ADAB and IRONOFF datasets, reveal promising results. In particular,

the comparison with the state of the art demonstrates the signiﬁcance of the proposed system.

1 INTRODUCTION

Graphs have emerged as an active area of research

aims to model structural relations of objects and pat-

terns. The graph’s ability to model different parts of

an object as well as its bases on sound mathemati-

cal background can be invested in many diverse ﬁelds

(Baldini et al., 2019; Lee et al., 2018). In the domain

of handwritten character recognition, graph drew the

attention and whetted the interest of numerous re-

searchers.

The use of a graph-based handwritten recognition

induces the need to formulate two main required oper-

ations: transforming handwritten graphs into feature

vectors and calculating the graph similarity. From this

perspective, the common task is to compare graphs to

ﬁnd the similarities between them. This is known as

graph matching (GM). Basically, two types of graph

matching were adopted by researchers(Yan et al.,

2016). The exact graph matching refers to the search

for an exact replication of the test graph in the tem-

plate graph as well as the conservation of all rela-

tionships presented in test one. The complexity of

the exact graph matching has not yet been speci-

ﬁed to be P or NP(Conte et al., 2004), but there are

https://orcid.org/0000-0002-7616-8374

polynomial algorithms for solving the isomorphism

problem of certain graph categories. A well-known

method is based on the depth-ﬁrst search (backtrack-

ing) with a forward checking method which greatly

reduces the number of backtracking steps. The in-

exact graph matching provides a distance value that

indicates graph dissimilarity(Bengoetxea, 2002). One

of the most ﬂexible and versatile approaches to inex-

act graph matching is graph edit distance(Abu-Aisheh

et al., 2017). However, the latter suffers from its high

complexity that limits its applicability to graphs with

small size. For this reason, a number of methods ad-

dressing the high computational complexity of graph

edit distance computation has been established, e.g.

(Darwiche et al., 2019).

Moreover, in recent years many tree-based

methods(Abu-Aisheh et al., 2015) have become of

great interest to researchers since computational time

and even the explored search space can be manage-

able with the impact of the quality of the provided

matching solution. Therefore, the primary motiva-

tion of the paper lies in tree-based methods which can

be explored in GM computation. Besides, owing to

the variability and ambiguity of handwritten charac-

ter strucure (for example: disorder, imprecision, con-

nection, etc) the use of fuzzy graph-based description

could be extremely helpful to add ﬂexibility against

Zitouni, R., Bezine, H. and Arous, N.

Graph based Method for Online Handwritten Character Recognition.

DOI: 10.5220/0008956602630270

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 1: GRAPP, pages

263-270

ISBN: 978-989-758-402-2; ISSN: 2184-4321

263

these errors. In this paper, we propose a novel fuzzy

graph-based handwritten method that takes advantage

of the aforementioned tree technique and fuzzy logic

concept merging them together with a Weighted Eu-

clidean Distance (WED) to matching two characters

graphs. Finally, we shall use the fuzzy attributed rela-

tional graph (FARG) matching algorithm to recognize

online hand written characters. The rest of this paper

is organized as follows. Section 2, presents a brief

description about related works. Section 3, illustrates

the graph used to represent the handwritten character

and presents the graph matching algorithm. In section

4, the experimental evaluation is elaborated on ADAB

and IRONOFF datasets. Finally, section 5 wraps up

the conclusion and provides new perspectives for fu-

ture works.

2 LITERATURE REVIEW

Great attention has been paid to graph based method

in many application ﬁelds. Among them, we can cite:

image analysis (Cl

ement et al., 2018), etc. For in-

stance, a few attempts are performed in online hand-

written character recognition research with graph rep-

resentation or relative concepts. One of the pioneer

attempts to use graph in online handwritten charac-

ter recognition was carried out by Si-Wei Lu et al.(Lu

et al., 1991). To obtain an efﬁcient recognition system

and avoid much of the combinatorial explosion in the

graph matching, the author resorted to a two-level hi-

erarchical graph representation for handwritten Chi-

nese characters. The two-level graph has a special

structure: the lower level is composed of nodes that

denote the strokes of a character to which they at-

tributed through the irrespective orientation on the

one hand. The edges denote relations between nodes

according to the type of junction on the other. Fi-

nally, an error-tolerant graph matching for classiﬁca-

tion framework is provided. This attempt is also con-

ducted with (Huang et al., 1993) to recognize Chi-

nese characters. Indeed, any character class will be

described by a graph where the stroke’s points repre-

sent the nodes and relation between them is deﬁned

by an edge. A constraint-based optimization prob-

lem is formulated using a relaxation matching for re-

solving issues in pattern recognition. Resting upon

Chinese characters, in(Suganthan and Yan, 1998) a

method for recognition of hand printed Chinese char-

acters is deﬁned. Hopﬁeld networks have been used

to solve graph matching problem. In both references

(Rocha and Pavlidis, 1994) and (Rocha and Pavlidis,

1995) the task of recognizing machine printed dis-

torted characters in text lines is tackled. Unlike all

graph methods each node corresponds to a branch,

an ending, a turning points or a sharp corner and the

edges represent strokes. Recognition task of text lines

turns into an error-tolerant subgraph isomorphisms

from character prototype graphs to text line graphs. In

this case, the beneﬁts recorded compared to feature-

based methods are typically two fold: First, mak-

ing optical character recognition to be independent of

segmentation. Second, handling both touching and

broken characters. Ref (Chakravarthy and Kompella,

2003) also tackled the online handwritten recogni-

tion problem using graph-based shape representation.

In this work, the authors aim to build graphs from

a handwritten shape. In this sense, they opted ﬁrst

for creating a vector out of seven Structural Proper-

ties (SP) related to the shape of handwriting: inte-

rior, end, bump, cross, dot and various other char-

acteristic shape points of the trace of a handwritten

text. Thereafter, each handwritten character is rep-

resented by a graph with SP design the nodes and

edge served as a link between nodes. (Al Mubarok

and Nugroho, 2016) proposed a method using the

skeleton structure of character as a way to design

graph where edges were identiﬁed into shape types

and nodes were detected with line simpliﬁcation al-

gorithm. The graph matching issue is converted into

hierarchical approach and is concepted with the sub-

graph isomorphism principals. To the knowledge of

the underlying authors, these methods are still far

away of being able to deal efﬁciently with a large and

multilingual dataset of complexe characters. Thus, we

evolve a fuzzy matching algorithm for graph-based

handwriting recognition. The proposed matching al-

gorithm permits efﬁciency more regarding the vari-

ability of characters, as well as the imprecision and

the ambiguity of their structures.

3 GRAPH REPRESENTATION OF

HANDWRITTEN CHARACTERS

3.1 Introduction of Fuzzy Attributed

Graph

The notion of Attributed relational graph (ARG) was

introduced in 1983 by Tsai and Fu for pattern analy-

sis task (Tsai and Fu, 1983). The nodes of the graph

indicate the different components of objects while the

edges denote the relations between these components.

More recently, this deﬁnition has been extended to a

Fuzzy Attributed Relational Graph (FARG) by attach-

ing fuzzy attributes to its nodes and edges. FARG was

introduced by Chang and Cheung (1992)(Chan and

GRAPP 2020 - 15th International Conference on Computer Graphics Theory and Applications

264

Cheung, 1992). These references (Jouili et al., 2009;

Luqman et al., 2013) are suggested for supplementary

reading on ARG and FARG respectively and their ap-

plications.

3.2 Graph Generation

An online character pattern is composed of a se-

quence of an arbitrary number of strokes. A stroke

is deﬁned as an elementary drawn between pen up

and pen down. A handwritten character is approxi-

mated by a graph where stroke is denoted with node.

These nodes are linked by an edge that represents the

relations between them. For a handwritten character

that is composed of n strokes, its associated graph is

also composed of n nodes (same number of strokes)

and n(n- 1)/2 edges. An example of nodes and edges

numbers is illustrated in ﬁgure 1.

Figure 1: (a) An example of a handwritten character, (b) Its

corresponding graph.

3.3 Computation of Node Attributes

Two nodes attributes are used namely type and size.

3.3.1 Stroke Type Attribute

Based on the βeta-elliptic model, each script in the

segmentation step is modeled in the dynamic domain

by a series of βeta proﬁles and in the static domain

by a series of elliptic arcs. The reader is referred to

(Bezine et al., 2007) for more details. Firstly, for

each elliptic stroke we assigned an Elementary Per-

ceptual Code (EPC) (Njah et al., 2012). However,

to solve the problems of perception and uncertainty

of assessed EPC, we opted for fuzzy logic representa-

tion(Njah et al., 2012). Based on the human vision we

cannot perceive these EPCs. What we can perceive is

restricted to some general forms labeled Global Per-

ceptual Codes (GPCs)obtained with the combination

of fuzzy EPCs. Due to the variety of possible com-

binations concerning either the number or the type of

EPCs that form a GPC, we opted to use the Genetic

Algorithms (GAs) to detect GPCs easily. Figure 2 ex-

hibits the different GPCs, their classiﬁcation and the

corresponding shapes.

Figure 2: List of GPCs.

3.3.2 Size Type Attribute

The following equation is used to compute the size of

handwritten stroke :

d =

− x

)

+ (y

− y

)

(1)

Where (x

) and (x

) denote respectively the start

and end point coordinates. The membership functions

for the linguistic values of this feature are identiﬁed

over the domain [0, 1]. The domain is then divided

into 5 fuzzy regions. In order to get more signiﬁcant

regions and minimum overlapping sets between them.

The linguistic labels associated with it are Very Small

(VS) , Small (S) , Medium (M), Big (B) and Very Big

(VB) as depicted in ﬁgure 3 .

Figure 3: Size representation.

3.4 Computation of Edge Attribute

3.4.1 Spatial Relations

Same strokes can appear differently in the same char-

acters and can be placed in different positions, which

may change the characters meaning. For this reason,

deﬁning the spatial relation is necessary. In spite of

the advantages, relationships have not been actively

modeled in many online handwritten character recog-

nition systems. They were tackled in many ofﬂine

task (Cl

ement et al., 2018). As a matter of fact, in or-

der to infer the relationship within handwriting stroke,

we should ﬁrst invest strokes’ topological properties

i.e centroid and bounding boxes.

Bounding Box: A stroke bounding box is deﬁned as

the smallest rectangle that ﬁrmly encloses the stroke.

Bounding box of stroke ’S’ is represented by four co-

ordinates: x

= min(x

), x

= max(x

), y

= min(y

)

and y

= max(y

) as depicted in the ﬁgure 4.

Centroid: The most common technique to check the

position of stroke within a region or not is to ana-

Graph based Method for Online Handwritten Character Recognition

265

Figure 4: Bounding box of an Arabic stroke.

lyze the coordinates of its centroid. We deﬁne the x-

coordinate of the centroid of a stroke s as x

= (x

−

)/2 and the y-coordinate as y

= (y

− y

)/2. At

this level, the consideration of two successive hand-

written strokes S and S

and the use of the center of

their bounding boxes and the centroids of both strokes

enable us to identify the following geometrical fea-

tures.

i) The angle of the line passes through the two cen-

troids of strokes with respect to the horizontal line

as expressed in the equation below:

θ = tan

−1

(

− y

− x

) (2)

ii) The intersection points

S ∩ S



1 if (S ∩ S

6= ∅) ⇒ intersect

0 otherwise

(3)

iii) The distance between the y-coordinates of the

centroids, denoted by dist

(S,S

) = y

− y

iv) The distance between the x-coordinates of the

centroids, denoted by dist

(S,S

) = x

− x

v) The distances between the x-coordinates

of the bottom right boundaries, denoted by

dist

(S,S

) = x

− x

vi) The distances between the x-coordinates of the

Top left boundaries, denoted by dist

top

(S,S

) =

− x

An illustrative example is presented in ﬁgure 5 with

the different calculated metrics.

Figure 5: Geometrical features between two handwritten

strokes S and S

of Latin character.

The different features serve as input to a fuzzy con-

troller whose outputs are a spatial relation type.

Therefore a set of ten spatial relations are identiﬁed

as illustrated in ﬁgure 6.

Figure 6: Strokes relationships.

3.5 Matching Handwritten Graphs and

Recognition

The comparison of two handwritten graphs can be for-

mulated as a problem of inexact graph matching. For

thus, we adopted an algorithm based a tree search of

the node mappings. The following sections describe

the main parts of this algorithm in detail.

3.5.1 Proposed Algorithm

Given two FARGs G

which denotes the template

handwritten graph and G

the test one. The graph

has N nodes, which are represented as n

(i ∈

1,2···N) while graph G

has N

nodes, which are rep-

resented as n

(i ∈ 1,2 · · ·N

). Assume that any n

can be expressed by a vector of µ

and µ

respec-

tively:

= µ

(i)

[1],µ

(i)

[2],...,µ

(i)

[a],...,µ

(i)

[A],a ∈ {1,2, ...A} (4)

= µ

(i)

[1],µ

(i)

[2],...,µ

(i)

],...,µ

(i)

[A],a

∈ {1,2,...A} (5)

A is the total number of membership degree associ-

ated with nodes regardless of n

. Different weights

are assigned to each node features according to their

relative importance to describe the node characteris-

tics. We denote by W

the weight of node n

and W

the weight of node n

. G

and G

are represented in

the weight space as a weight vector of W

and W

respectively :

,...,W

,i ∈ {1,2,...N}and

∑

i=1

= 1 (6)

,...,W

,i ∈ {1,2,...N

}and

∑

i=1

= 1 (7)

node n

and node n

with the highest value of W

and W

respectively are considered the most impor-

tant nodes during the nodes matching.

The edge between every node n

and n

in G

is rep-

resented by a vector of e

(i, j)

while the edge between

every node n

and n

in G

is represented by a vec-

tor of e

(i, j)

. Similarly, both edges e

(i, j)

and e

(i, j)

are

represented by a set of vectors denoted below :

(i, j)

= e

(i, j)

[1],e

(i, j)

[2],..., e

(i, j)

[b],..., e

(i, j)

[B],b ∈ {1, 2, ...B} (8)

(i, j)

= e

(i, j)

[1],e

(i, j)

[2],..., e

(i, j)

],..., e

(i, j)

[B],b

∈ {1, 2, ...B} (9)

GRAPP 2020 - 15th International Conference on Computer Graphics Theory and Applications

266

Where B is the total number of membership associ-

ated with edges regardless of e

(i, j)

or e

(i, j)

respec-

tively. If we denote by W

the weight of edge e

(i, j)

and W

the weight of edge e

(i, j)

. G

and G

are rep-

resented in the weight space as a weight vector W

and W

respectively :

,...,W

,i ∈ {1,2,...T }and

∑

i=1

= 1 (10)

,...,W

,i ∈ {1,2, ...T

}and

∑

i=1

= 1 (11)

With T and T

are the total number of edges in tem-

plate and test character respectively.

The graph matching is carried out by means of a tree

search algorithm provided in algorithm 1.

Algorithm 1: Tree-based search algorithm.

Input: Template graph G

and test graph G

Output: The lowest cost match between G

and G

Matching(G

)

Dist

min

= ∞

P =

min

=NULL

N’=

N ← extract

nodes

)

N’ ← extract

nodes

)

for each node n ∈ N do

for each node n

∈ N

Compute Dist(n,n

) using Eq(10)

if (Dist(n,n

) ≺ Dist

min

) then

Dist

min

=Dist(n,n

)

min

end if

end for

P ← P ∪ {(n,N

min

)}

N ← N - {(n)}

← N

- {(N

min

)}

end for

Return P

In order to identify the similarity between two entities

a distance metric is needed. From this perspective,

the weighted euclidean distance metrics opts for the

fuzzy attributes. Given a node n

in G

and its pair

node in G

is n

, a distance between the two nodes is

measured using weighted Euclidean metric (WED) as

portrayed below:

Dist

= [

∑

i=1

× (µ

− µ

)

]

(12)

The weights W

are proposed as:

= max(max(µ

),max(µ

)) (13)

It is noticeable that a distance between two paired

edges is computed by the equation below:

Dist

= [

∑

i=1

× (e

(i, j)

− e

(i, j)

)

]

(14)

The weights W

are proposed as:

= max(max(e

(i, j)

),e

(i, j)

)) (15)

A synthesis distance caused by all nodes is then ex-

pressed by:

Dist

nodes

∑

i=1

Dist

(16)

A synthesis distance caused by all edges is then ex-

pressed by:

Dist

edges

∑

i=1

Dist

(17)

Where L

and L

are the number of matched node

pairs and the number of matched edge pairs, respec-

tively. Thus a similarity measure between a pair of

FARG r can be deﬁned directly with

Sim(G

) = W

× Dist

nodes

× Dist

edges

(18)

Where W

and W

are properly selected weights of

and G

. The procedure of computation of similar-

ity is summarized below (algorithm 2).

Algorithm 2: The proposed FARG matching algorithm.

Input: Template graph G

and test graph G

Output: Matching similarity measure Sim

Dist

nodes

Dist

edges

=matching( G

, G

)

for each successor pairs of nodes {(n,n

),(n

)} ∈

P1 do

=extract

edge

(n,n

) ∈ G

=extract

edge

) ∈ G

Dist

=Compute Dist(E

) using Eq(12)

Dist

edges

← Dist

edges

∪ {(Dist

)}

end for

for each pair of nodes (n,n

) ∈ P1 do

Dist

=Compute Dist(n,n

) using Eq(10)

Dist

nodes

← Dist

nodes

∪ {(Dist

)}

end for

Dist

nodes

← Dist

nodes

∪ {(Dist

)}

Compute Sim using Eq(16)

3.6 Recognition

The proposed approach uses a fuzzy graph matching

algorithm to recognize online handwritten characters.

Graph based Method for Online Handwritten Character Recognition

267

Figure 7: Flowchart of the proposed algorithm for hand-

written recognition characters.

Figure7 illustrates an overview of the framework and

rests upon a set of phases as described below.

In the ﬁrst phase, the FARG of handwritten characters

for both template and test are extracted as explained in

section 3. In the second phase the similarity of nodes

Sim(nodes)

i j

as well as Sim(edges)

i j

between a graph

FARG

of the i

test graph and a graph FARG

of the

template graph is calculated as mentioned in al-

gorithm 2. Then the similarity S

between the two

graphs is computed by summing up the nodes and

edges similarities. Therefore, the class correspond-

ing to the graph which provides the best match (max-

imum degree of similarity) constituent the recognized

handwritten character.

C = argmax

{Sim

} (19)

4 EXPERIMENTAL RESULTS

At this stage of analysis, we shall address the obtained

results, discuss them and evaluate them in order to as-

sess the effectiveness and feasibility of our approach.

4.1 Datasets

ADAB Data Base: The data base ADAB (El Abed

et al., 2011) was developed in cooperation between

the Institut fuer Nachrichtentechnik (IfN) and Re-

search Groups in Intelligent Machines,Tunisia at Uni-

versity of Sfax. It was collected by using digital

tablets connected to computers,more than 130 dif-

ferent writers mainly of Tunisian nationality. This

database contains 21575 online Tunisian town and

village names in Arabic languages. IRONOFF

database: The database IRONOFF (Viard-Gaudin

et al., 1999) is created using the mechanism Pen

Tablet which was of a tablet type A4Wacom Ultra Pad

sampling speed of 100 points per second since 1999.

The prototype is recorded in the format UNIPEN.

IRONOFF database is characterized by the domi-

nance of isolated symbols, French and English words

as well as digits.

A set of 1135 handwritten characters are randomly

selected from these data bases to perform fuzzy graph

matching and therefore to carry out the recognition

task. First, each handwritten words is converted into

graphs by representing strokes by nodes and spatial

relationships between them by edges. Table 1 por-

trays the graph and its distribution degree. We denote

the ex, tr and tes as the size of examples, training and

testing of both datasets respectively. Minimum and

maximum of nodes N and edge E degrees for graphs

are also presented. The size varies from small graphs

with under 5 nodes and edges and up to 15 nodes and

edges in both datasets.

Table 1: Graph and their degree distribution.

Size Min Max

Dataset ex tr tes |N| |E| |N| |E|

IRONOFF 634 381 253 2 1 10 45

ADAB 519 312 207 2 1 15 105

4.1.1 Experimental Results

In order to explain why the FARG has such an ex-

cellent performance, time complexity should be ana-

lyzed.

4.1.2 Execution Time

The different experiments were performed on PC with

Intel Core i7-7500, 8 GB RAM. Table 2 displays the

required time of each input character for generating

(building) its corresponding graph related to the nodes

number.

Table 2: Computational time for graph generation.

Number of nodes Graph generation times

≺ 5 ≺ 10s

5-10 10-25 s

 10  25s

It can be observed that the method requires less ex-

ecution time for a smaller number of nodes. Ta-

ble 3 introduces the running time for some hand-

written drawing. For each entry we would illustrate

the drawn characters property number of nodes of

template and test graph (N

, N

, respectively) ,the

time for generating both graphs(GG) N

and N

the time for graph matching between graphs(M) (sec-

tion 3) and ﬁnally the total computational time. Note

that the largest computational times are required for

GRAPP 2020 - 15th International Conference on Computer Graphics Theory and Applications

268

generation (building) graphs mainly due to the pre-

processing step which was performed to reduce the

noise and step needed for calculating nodes/edges at-

tributes.

Table 3: Computational time.

Scripts |N

| |N

| Computational time

GG M Total

”



é¢ËA



®J

Ë@” 11 11 56 22 78

”ÉêÓñK

” 4 4 18 7 25

ñJ









” 7 7 30 9 39

”



éÓ



” 6 6 24 8 32

”



èð@Q



ªÓ” 7 7 40 17 57

”





KAK





@” 10 10 46 18 64

”A” 3 3 14 6 20

”B” 2 2 12 4 16

”Money” 2 2 14 5 19

”Quiz” 5 5 18 6 24

”quelqu’un” 8 8 42 19 61

”Neptune” 8 8 48 15 63

4.2 Recognition Accuracy

For both dataset, the training and testing sets contain

60% and 40% of graphs respectively, and serve as the

dataset graphs. Referring to table 4 and table 5, we get

an impressive recognition rate which exceeds 98% for

both dataset. Next, our ultimate objective is to eval-

uate our approach which is based on graph matching

with related works. It is worth noting, a large vari-

ety of recognition algorithms has been tested on these

datasets and some of the best results are outlined.

Table 4: Best reported recognition results for the ADAB

dataset.

Systems Recognition rate

HMM (Khlif et al., 2016) 93.33%

MLP-HMM (Elleuch et al., 2016) 96.45%

CNN (Tagougui et al., 2013) 91.8%

AUC-HMM1 (El Abed et al., 2011) 98.45%

FARG matching (present work) 98.62 %

In table 4, we compared our results with re-

sults conducted on ADAB dataset by (Khlif et al.,

2016),(Elleuch et al., 2016), (Tagougui et al., 2013)

and (El Abed et al., 2011) and based on HMM, MLP-

HMM, CNN networks and AUC-HMM1 respectively.

The recognition rate provided by our algorithm ex-

hibits an improvement that varies between 6.82% and

0.17% over the four participating systems.

Table 5 reports also the performance of our recogni-

tion systems on IRONOFF dataset. We achieved a

Table 5: Best reported recognition results for the IRONOFF

dataset.

Systems Recognition rate

ConvLSTM(Akouaydi et al., 2019) 93%

BCP+LSTM(Akouaydi et al., 2019) 98%

FARG matching (present work) 99.24%

very competitive rate with respect to other two pub-

lished result (Akouaydi et al., 2019) based on Con-

vLSTM and BCP+LSTM. We exhibited an improve-

ment that varies between 6.24% and 1.24% over re-

lated systems. This ﬁnding which based on the use

of an efﬁcient WED metric jointly with a tree search

process guides the matching solutions and therefore

contributes to a high-rate recognition.

5 CONCLUSIONS

In this paper, we aimed at setting forward a new ap-

proach to perform an online handwritten recognition.

The nodes and edges of the graph are labeled using

fuzzy logic attributes. The graphs are matched us-

ing a WED as the similarity function and based on a

tree search process. The approach is tested on ADAB

and IRONOFF datasets. This approach helps to rec-

ognize the handwritten character and achieve recog-

nition rates of 98.62 % and 99.24% respectively. Our

work is an initial step that can be further developed

as it opens new perspectives and offers different hori-

zons for future works in the domain of handwritten

recognition includes the investigation of a graph em-

bedding method with a new classiﬁcation schemes

such as deep learning.

REFERENCES

Abu-Aisheh, Z., Ga

uzere, B., Bougleux, S., Ramel, J.-Y.,

Brun, L., Raveaux, R., H

eroux, P., and Adam, S.

(2017). Graph edit distance contest: Results and fu-

ture challenges. Pattern Recognition Letters, 100:96–

103.

Abu-Aisheh, Z., Raveaux, R., Ramel, J.-Y., and Martineau,

P. (2015). An exact graph edit distance algorithm for

solving pattern recognition problems. In 4th Inter-

national Conference on Pattern Recognition Applica-

tions and Methods 2015.

Akouaydi, H., Njah, S., Ouarda, W., Samet, A., Dhieb, T.,

Zaied, M., and Alimi, A. M. (2019). Neural archi-

tecture based on fuzzy perceptual representation for

online multilingual handwriting recognition. arXiv

preprint arXiv:1908.00634.

Al Mubarok, A. and Nugroho, H. (2016). Handwritten char-

acter recognition using hierarchical graph matching.

In International Conference on Advanced Computer

Graph based Method for Online Handwritten Character Recognition

269

Science and Information Systems (ICACSIS), pages

454–459. IEEE.

Baldini, L., Martino, A., and Rizzi, A. (2019). Stochastic

information granules extraction for graph embedding

and classiﬁcation. In Proceedings of the 11th Inter-

national Joint Conference on Computational Intelli-

gence, volume 1, pages 391–402.

Bengoetxea, E. (2002). Inexact Graph Matching Using Es-

timation of Distribution Algorithms. PhD thesis, Ecole

Nationale Sup

erieure des T

ecommunications,Paris,

France.

Bezine, H., Keﬁ, M., and Alimi, A. M. (2007). On the beta-

elliptic model for the control of the human arm move-

ment. International Journal of Pattern Recognition

and Artiﬁcial Intelligence, 21(01):5–19.

Chakravarthy, V. S. and Kompella, B. (2003). The shape

of handwritten characters. Pattern recognition letters,

24(12):1901–1913.

Chan, K. and Cheung, Y. (1992). Fuzzy-attribute graph

with application to chinese character recognition.

IEEE transactions on systems, man, and cybernetics,

22(1):153–160.

ement, M., Kurtz, C., and Wendling, L. (2018). Learning

spatial relations and shapes for structural object de-

scription and scene recognition. Pattern Recognition,

84:197–210.

Conte, D., Foggia, P., Sansone, C., and Vento, M. (2004).

Thirty years of graph matching in pattern recognition.

International journal of pattern recognition and arti-

ﬁcial intelligence, 18(03):265–298.

Darwiche, M., Conte, D., Raveaux, R., and T’Kindt, V.

(2019). A local branching heuristic for solving a graph

edit distance problem. Computers & Operations Re-

search, 106:225–235.

El Abed, H., Kherallah, M., M

argner, V., and Alimi, A. M.

(2011). On-line arabic handwriting recognition com-

petition. International Journal on Document Analysis

and Recognition (IJDAR), 14(1):15–23.

Elleuch, M., Zouari, R., and Kherallah, M. (2016). Feature

extractor based deep method to enhance online arabic

handwritten recognition system. In International Con-

ference on Artiﬁcial Neural Networks, pages 136–144.

Springer.

Huang, X., Gu, J., and Wu, Y. (1993). A constrained

approach to multifont chinese character recognition.

IEEE transactions on pattern analysis and machine

intelligence, 15(8):838–843.

Jouili, S., Mili, I., and Tabbone, S. (2009). Attributed graph

matching using local descriptions. In International

Conference on Advanced Concepts for Intelligent Vi-

sion Systems, pages 89–99. Springer.

Khlif, H., Prum, S., Kessentini, Y., Kanoun, S., and Ogier,

J.-M. (2016). Fusion of explicit segmentation based

system and segmentation-free based system for on-

line arabic handwritten word recognition. In 15th In-

ternational Conference on Frontiers in Handwriting

Recognition (ICFHR), pages 399–404. IEEE.

Lee, J. B., Rossi, R. A., Kim, S., Ahmed, N. K., and Koh, E.

(2018). Attention models in graphs: A survey. arXiv

preprint arXiv:1807.07984.

Lu, S. W., Ren, Y., and Suen, C. Y. (1991). Hierarchi-

cal attributed graph representation and recognition of

handwritten chinese characters. Pattern Recognition,

24(7):617–632.

Luqman, M. M., Ramel, J.-Y., Llad

os, J., and Brouard, T.

(2013). Fuzzy multilevel graph embedding. Pattern

Recognition, 46(2):551–565.

Njah, S., Ltaief, M., Bezine, H., and Alimi, A. M. (2012).

The pertohs theory for on-line handwriting segmenta-

tion. International Journal of Computer Science Is-

sues (IJCSI), 9(5):142–151.

Rocha, J. and Pavlidis, T. (1994). A shape analysis model

with applications to a character recognition system.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 16(4):393–404.

Rocha, J. and Pavlidis, T. (1995). Character recognition

without segmentation. IEEE Transactions on pattern

analysis and machine intelligence, 17(9):903–909.

Suganthan, P. N. and Yan, H. (1998). Recognition of

handprinted chinese characters by constrained graph

matching. Image and Vision Computing, 16(3):191–

201.

Tagougui, N., Boubaker, H., Kherallah, M., and Alimi,

A. M. (2013). A hybrid mlpnn/hmm recognition sys-

tem for online arabic handwritten script. In World

Congress on Computer and Information Technology

(WCCIT), pages 1–6. IEEE.

Tsai, W.-H. and Fu, K.-S. (1983). Subgraph error-

correcting isomorphisms for syntactic pattern recog-

nition. IEEE Transactions on Systems, man, and cy-

bernetics, (1):48–62.

Viard-Gaudin, C., Lallican, P. M., Knerr, S., and Binter,

P. (1999). The ireste on/off (ironoff) dual handwrit-

ing database. In Proceedings of the Fifth Interna-

tional Conference on Document Analysis and Recog-

nition(ICDAR), pages 455–458. IEEE.

Yan, J., Yin, X.-C., Lin, W., Deng, C., Zha, H., and Yang,

X. (2016). A short survey of recent advances in graph

matching. In Proceedings of International Conference

on Multimedia Retrieval, pages 167–174. ACM.

GRAPP 2020 - 15th International Conference on Computer Graphics Theory and Applications

270