You are on page 1of 12

Auto-generated PDF by ReView

This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

IET Research Journals

Brief Paper

Graphology based Handwritten Character Analysis for Human Behavior Identification

Subhankar Ghosha, Palaiahnakote Shivakumarab, Prasun Roya, Umapada Pala and Tong Luc

a
Indian Statistical Institute, Kolkata,700108, India
b
Faculty of Computer Science and Information Technology, University of Malaya, 50603, Kuala Lumpur, Malaysia
cNational Key Lab for Novel Software Technology, Nanjing University, 210093, Nanjing, China

Abstract: Graphology based handwriting analysis to identify behaviors of individuals, irrespective of applications and fields
without human intervention, is interesting and useful. Unlike existing methods that use characters, words and sentences for
behavioral analysis, which often require trained individuals and are pertaining to specific applications, we propose an automatic
method by analyzing a few handwritten English lowercase characters from a to z to identify person behaviors. The proposed
method extracts structural features, such as loops, slants, cursive, straight lines, stroke thickness, contour shapes, height, width,
aspect ratio and other geometrical properties, from different zones of isolated character images to derive the hypothesis based
on dictionary of Graphological rules. The derived hypothesis has the ability to categorize personal, positive, and negative social
aspects of an individual. To evaluate the proposed method, we developed an automatic system which accepts characters from
a to z written by different individuals across different genders as well as different age groups. This automatic privacy projected
system is available in the website (http://subha.pythonanywhere.com). For quantitative evaluation of the proposed method,
several people are requested to use the system to check their characteristics with the system automatic response based on
his/her handwriting by choosing agree or disagree options. So far, we have received 5300 responses from the users, on which
we noted that the accuracy of our system is 86.70%.

KEYWORDS: Human reading, Human Behavioral Identification, Graphology, Handwriting recognition.

1. INTRODUCTION considers a part of a character. Similarly, there are studies on


aesthetic analysis [5,6], which use handwritten characters or
Graphology based behavioral analysis is gaining
image features to predict personal behaviors such as beautiful,
popularity in the recent years due to widespread applications
non-beautiful, excellent writing and poor writing. However,
across diverse fields, such as psychology, education, medicine,
these methods are limited to few behavior types. There are
criminal detection, marriage guidance, commerce and
methods proposed for personality assessment [7,8], which
recruitment [1] etc. In addition, it is also noted that graphology
assess person interest, attitude, relationship with family,
based handwriting reveals inner feelings of persons though
community, etc. These methods study typed texts by users but
such characteristics are invisible from person behaviors [2].
not handwritten texts. In general, the methods require full text
Therefore, traditional methods that use visible facial/biometric
lines or sentences for personality assessment. Therefore, to
features or human actions to identify person behaviors may not
predict or identify personal behaviors which may be invisible
be effective, especially when a person pretends artificially.
from eyes, graphology based handwriting analysis becomes
Besides, most conventional methods are application, situation
essential.
and dataset dependent. Thus, graphology based handwriting
analysis is used as an objective tool for studying person It is noted that graphology based human behavior analysis
behaviors without depending on appearance based features of does not have mathematical basis. However, the rules and
persons to make a system independent on fields, data, gender, definition are derived based on the discussions and views of
age of a person, applications, etc. Further, since graphology the people of a graphological institute in Kolkata (http://mbose-
focuses on individual letters, strokes and part of a character kig.com/, India). Also some of the rules are obtained based on
rather than the whole character, word, or document, features the experience and psychology of the people as discussed in
will be sensitive to personal behaviors, which help in the link (https://ipip.ori.org/). It is evident from the literature
predicting person behaviors [1]. Several methods have been [7, 8], where we can see several methods are published for
proposed for predicting person behaviors using graphology studying the human behavior using graphology. For example,
based handwriting in literature [1,2]; however, such methods when the person is under pressure and in poor condition, it is
expect human intervention to identify behaviors. Therefore, expected such behavior reflects in writing. Therefore, the
there is an urgent need for developing a generalized shape of character changes compared to writing when the
graphology based handwriting analysis method. To study person is normal. The same changes are extracted using
behaviors such as emotions, feelings and person features for human behavior identification in this work. This
identifications, there are methods [3,4] which use signature for is the basis that we used for proposing the method for human
identification. However, these methods consider the whole behavior identification using handwriting analysis.
signature for prediction in contrast to graphology which
IET Research Journals, pp. 1–11
©The Institution of Engineering and Technology 2019 1

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 2


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

2. RELATED WORK explores generalized Hough transform for studying the


orientation of character “y” is either left, right vertical or
To the best of our knowledge, many methods have been
extreme right skewed. Comparing to [11], this work focuses
proposed on graphology based handwriting in literature [1,2].
on the attributes of character “y”. Therefore, the scope is
However, most of the methods describe theory, concepts and
limited to specific characters but not a generalized method.
usefulness. As a result, developing automatic systems for
personal behavior identification based on graphology Coll et al. [13] proposed graphological analysis of
handwriting analysis is a new study in document analysis. handwritten text documents for human resource recruitment.
In this work, the method focuses on attributes, such as active
Asra and Shubhangi [8] proposed human behavor
personality and leadership, which are required for human
recognition based on handwriten cursive by an SVM classifier.
resource recruitment. The features such as layout
The method extracts geometrical features such as shape, stroke
configuration, letter size, shape, slant and skew angle of lines,
and corner information for segmeted regions of interest, and
are considered for the above personal trait identification. The
then the features are passed to an SVM classifier for human
method explores projection profile based features, contour
behavior recognition. However, types of person behaviors
based features, Discrete Cosine Transform and entropy for
they consider for recognition are not mentioned. In addition,
extracting the above features from handwritten documents. It
the method consdiers regions of interest for feature extraction.
is noted that the method requires the full document and focuses
As a result, if the method does not get correct regions of
only on the requirement of human resource recruitment.
interest, it may not perform well. In addtion, the features
Therefore, their scope is limited to specific applications.
extracted are sensitive to background and foreground.
In summary, the methods [11-13] focus on extracting
Fallah and Khotanalou [9] proposed to identify human
features such as writing force, pressure as well as shapes of
personlity parameters based on handwriting and neural
characters. Besides, the scopes of the methods are limited to
networks. The method extracts wirting style to find personality
specific behaviors and particular characters but not general
parameters of a person. In other words, the method finds
behaviors with multiple characters as the proposed work. This
variations in writing using correlation estimation and feature
is because the methods are proposed for particular applications.
extraction. Extracted features are fed to a neural network
Further, the features extracted based on writing force and
classifier for personality parameters detection. However, it is
pressure may not be effective compared to shape based
not clear about the number of parameters and the basis for
features as one can expect the same from different persons. In
parameter selection. In additon, the method requires at least
addition, the pressure and force depend on paper quality,
one word written by a person for persoanlity parameter
thickness, pen, ink, etc. On the other hand, in contrast to the
deteciton. Though the above two methods are related to person
existing methods [11-13], the proposed system does not target
behavior identification, the scopes and the ways they extract
any particular application and is not limited to specific
features are differrent from the proposed method.
characters. As a result, the way the proposed method extracts
Topaloglu & Ekmekci [10] proposed gender detection for features and defines rules based on graphologist is different
handwriting analysis. The method extracts attributes of from the above-mentioned existing methods. The key
handwritten characters for identifying gender, male or female. advantage of the proposed system is that it is independent of
The method works based on the fact that texts written by a application, character, age, gender, ink, paper, pen, etc.
female is often neat, visible and legible, and one can expect Therefore, we can argue that the proposed system tends to
uniform spacing between words, text lines, etc. While from generalization compared to the existing methods [11-13].
male writing, it is hard to find the above characteristics.
In light of the above discussions, it is observed that most
Therefore, the method extracts pressure, border, space,
of the methods are developed for specific applications. In
dimension of base lines, slanting, etc. In total, the method
addition, none of the method studies handwritten characters
extracts 133 attributes and then uses a decision tree for
from “a” to “z” for identifying possible personal behaviors. As
classification. Though the method studies graphology using
a result, we can conclude that there is no method which works
handwritten characters, the extracted attributes are limited to
well irrespective of applications, genders and fields. Therefore,
two classes. In addition, their scope is gender classification but
in this work, we propose to study attributes of characters from
not personal behavior identification.
“a” to “z” for the identification of personal behaviors. The
Champa & Kumar [11] proposed artificial neural contributions are as follows. (1) Exploring local information
networks for human behavior prediction through handwriting of handwritten characters for personal behavior identification
analysis. The method extracts baseline with its inclination and without depending on applications, fields, genders and ages,
writing pressure to identify personal behaviors such as (2) Deriving hypothesis based on local information and
different levels of emotions and confidence. For this purpose, dictionary of graphology for each behavior of persons, and (3)
the method uses character “t”. In other words, the method is An automatic interactive system for generating ground truth
limited to study the attributes of character “t”. However, the and validating the proposed method.
scope of the proposed work is to study the attributes of
characters to identify different types of personal behaviors. 3. PROPOSED METHOD
Champa & Kumar [12] also proposed automated human It is evident from graphology theory (https://ipip.ori.org/)
behavior prediction through handwriting analysis. The above that personal behaviors are often reflected in their handwriting
authors developed a method for identifying personal traits styles, especially in written loops, stems, height, width and
using handwriting character analysis. In this work, the method
IET Research Journals, pp. 1–11
2 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 3


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

slant of characters with respect to baseline [14]. It is also true the same way, for character “h”, the concavity created in the
that since our intention is to collect datasets online using pen bottom and the hole created at the top help us to find zones.
and pad, characters written by different users do not pose any For character like “c”, since there are no intersection points
distortion, degradation and noise. This observation motivates and branches, the concavity determines the middle zone. For
us to propose features, such as zone, angle and loop that extract loop based features, while tracing a boundary, if the proposed
effects and reflections of different behaviors of persons. The method visits the same starting point again, it can be
rationale behind to propose rules based features is as follows. considered as a loop or a hole. For the purpose of finding holes
When images are clean and preserve unique shapes with and loops, we have explored Euler number concept as shown
respect to different behaviors of persons, and graphology in Fig. 2(b). For angle based features, the proposed method
provides rules for each behavior of persons, we believe the uses curvature concept, which estimates angle for the region
proposed features with rules can achieve better results. The that have corners,where contour is cursive as shown in Fig.
proposed rules are used to study shapes and positions of loops, 2(c). For finding ends, intersections and branches of characters,
height or angle of stems with respect to baseline, end or branch the proposed method uses neighborhood information along
points, sleeping lines over characters, etc. with angle information as shown in Fig. 2(d) and Fig. 2(f). For
Note that the aim of the proposed work is to study personal extracting oval shaped features, the proposed method then
behaviors (e.g., talkative, broadminded, etc.), positive social explores the well-known image processing concept called
(e.g., witty, the ability to make work successful, etc.), negative water reservoir model [14], which defines oval based on water
social behaviors which create a bad social behaviors which collection in the oval area as shown in Fig. 2(e). Similarly, the
create cool social environment (e.g., irritating, cheating, etc.), proposed method explores run-length smearing concept,
and personal behavior which describes individual person which counts successive pixels of the same information for bar
attitude (e.g., ability, self-confidence, nervousness etc).To detection while tracing the boundary of a character as shown
achieve the above-mentioned goal, the proposed method in Fig. 2(g). Overall, the proposed method uses the above-
considers handwritten characters written by different persons mentioned image processing concepts for feature extraction in
as the input. We propose to extract structural features such as this work. Similarly, the proposed method extracts the
Zone based, Loop based, Angular based and other geometrical observations based on loop shapes by estimating height, width
information of characters for identifying personal behaviors. as shown in Fig. 2(b). These observations are useful for
According to graphologists and with their experiences, we identifying negative social behaviors. The proposed method
then derive hypothesis using structural features for identifying extracts observations based angle information as shown in Fig.
three types of person behaviors, namely, Positive social 2(c), which is useful to identify personal behaviors such as
behavior, Negative social behavior and Personal behavior. The “the ability to handle critical situations by signing for himself”.
flow of the proposed method can be seen in Fig. 1. Observations are extracted based on height and width of stems
Handwritten characters of different persons

Structural features extraction

Zone Loop Angle Other

Hypothesis for person behavior identification

Positive social behavior Negative social behavior Personal behavior

Fig. 1. Block diagram of the proposed method


with respect to baseline as shown in Fig. 2(d), which are useful
for the three types of behaviors.
3.1 Structural Feature Extraction
In this work, as mentioned in the previous section, we Observations based on oval shapes are extracted as shown
extracts different types of structural features, namely, Zone in Fig. 2(e), where the proposed method uses water reservoir
based, Loop based, Angular based, Stem based, Oval based, model [14] for identifying different types of ovals. These
Branch based, Bar based etc, for human behavior observations are useful for all the three types of personal
identification. For extracting Zone based features, the behaviors. Observations on end and branch points are
proposed method traces contours of characters to detect extracted using character skeletons as shown in Fig. 2(f),
intersection points, holes, concavity and convex hull, which which are also useful for identifying the three types of personal
are image processing concepts. Based on geometrical behaviors. Note that the proposed method extracts observation
properties of the above shapes, the proposed method divides based on bar (sleeping line over the characters) as shown in
each whole character into three zones, namely, upper, middle Fig.2(g), which are useful for identifying negative social
and lower zones as shown in Fig. 2(a). In Fig. 2(a), it is noted behavior of the person.
from character “y” that the intersection point and the hole are
helping the method to find lower, middle and upper zones. In
IET Research Journals, pp. 1–11
©The Institution of Engineering and Technology 2019 3

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 4


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

(a) Zone based features

(b) Loop based features

(c) Angular based features (d) Stem based features

(e) Oval based features (f) End point and branch point based features

(g) Bar based features

Fig. 2. Different structural features for deriving hypothesis

IET Research Journals, pp. 1–11


4 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 5


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

In the same way, we study the characteristics of the above Width of lower
structures of the different characters from “a” to “z” to identify 9 d
loop should be
Restless
less than its
unique observations for different personal behavior which are
height.
quite common and essential to make system successful in
respective fields such as education, medical management etc. Stem with long
Sudden Angry
The unique observations and respective behavior of the person loop 45 angle
10 d and cool down
deviation from
with handwritten characters are listed in Table 1, Table 4 and fast
baseline
Table 6, respectively for personal behavior, positive social and
negative social behavior. To extract the observations from End part of “e”
Table 1, Table 4 and Table 6, the proposed method derives 11 e continues with Attention seeker
hypothesis and implement the same to identify the behavior 45 degree angle
automatically as conditions are listed in Table 3, Table 5 and Good in
Starting and end
Table 7 for respective observations listed in Table 1, Table 4 12 f point of loop at
planning but
and Table 6. The acronym and meaning the variables are listed lack of
left side
in Table 2 which defines all the variables listed in Table 3, confidence
Table 5 and Table 7.
Narrow loop Intelligent and
Table 1. List of hypotheses derived based structural features of 13 h
with long stem open minded
handwritten character images for personal behavior.
Handwritten
NO. Letter Observations Behavior
Character
Loop Wide loop with Intelligent and
14 h
1 a disconnected at Talkative long stem broad minded
right.
Upper loop
appear as Loop of Balloon
Intelligent but
balloon and 15 h shape with long
2 b Broadminded have confusion
lower loop has stem
more width than
He loves
its height.
Spreading legs freedom and
The upper loop 16 h
of the character does not like to
is narrow and
Follow Idealism be controlled
3 b long and there is
and Talkative
small gap in
lower loop.
17 h Knot at top Frustration

Curve within
Hiding his/she
b, p, lower loop. Ring Cap like a shape
4 Egoistic 18 i own
m like structure at on top
weaknesses
the left of “m”

Dot at upper-
Advance
19 i right side of
thinker
character
Socially
5 c Semi-circle
comfortable
Highly stressful
20 i There is no dot
personality
Extra like
6 c tangent to Stubborn
Left hump is
contour
21 m higher than right Diplomatic
hump
Right hump is
Mirror image of 22 m greater than left Self-conscious
7 c Introvert
“C” hump

Disconnection in
Upper loop can 23 P Talkative
loop
be any sized and
Lazy (Wasting
width of lower
8 d time by
loop must be
watching TV)
higher than
height

IET Research Journals, pp. 1–11


2 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 6


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

L = Loop LH = Loop Heigh LW = Loop Width


End point is
Aggressive in
24 q stretched with San = Stem Angle with
profession
angle towards up S = Stem Baseline
Sh = Stem Height

End point O = Oval OH = Oval Height OW = Oval Width


25 q stretches like a Stressful
hook WR = NWR B = Number of NWR T = Number of
Water water Reservoir water Reservoir
reservoir From Bottom From Top
No stretching at
26 q Loner
end point
WR H NWR R = Number of
WR W = Width of
= Height of water Reservoir
Water Reservoir
Water Reservoir From Right
Ability to
27 r Angular
handle/manage
HM = LHMh = Left Hump RHMh = Right Hump
Hump of m Height of m Height of m
Letter looks like
Sing for herself PB = Branch Point
28 r “U” with loop at
or himself P = Point PE = End Point
left side

Loop end with CRB = Curb CRBH = Curb Height CRBW = Curb Width
Stress and
29 s stretch towards
Stubborn
right ZU = Upper Zone ZM = Middle Zone
Z = Zone
ZL = Lower Zone
Step and bar are
30 t Rigid
almost straight B = Bar Bl = Bar Left point Br = Bar Right point

Bw = Bar Width Bh = Bar Height


Starting or
31 v ending strokes Glamorous ∆x = X Cordinateof the ∆y = Y Cordinate of the
look line. ∆= Point
point point

EPe = right
Sw = Stroke Width Id = I Dot
most endpoint
Regular V, Analytical of e
32 v, z
Regular Z Person BPe = Lower
H = Hat
most point of e Line = Straight Line

Left and right Table 3. Conditions for the hypotheses listed in Table 1.
strokes are
33 i, Independent N 1 H = 0 &∆x (Id )
perpendicular to
Equations for the Hypotheses
base line. 0 9 < 𝑀𝐴𝑋(∆x (S))
2
1 NWR T = 1 Id = 0&𝐻 = 0
0
Loop end 2
Rigid, 2 LH < 3 ∗ LW &OH < OW LHMh > 𝑅𝐻Mh
stretches towards 1
34 y frustration and
right like straight 2
fear
line 3 NWR T = 0 &LH < 2 ∗ LW LHMh < 𝑅𝐻Mh
2
NWR B = 3 for m
2
4 NWR B = 2 for b NWR T = 1
Two Wavy bars 3
NWR B = 3 for p
on top and 2
bottom for “I”. 5 WR W ≅ 2 ∗ WR H ∆y (O) < ∆y (Epoint )
Fun loving 4
35 i, t, Z Wavy line for
person 2
“t”. Wavy stroke 6 NWR T = 1 H=1
at the end for 5
“Z". 7 2
NWR T = 1 H = 0& 𝑂 = 1&𝐿 = 0
6
2
8 OH < OW Apoint = 0
7
2
9 OH > Ow Apoint = 1
8
Table 2. Acronyms used for hypotheses derivation listed in Table 3,
1 2 ∆x (Epoint(2) )
Table 5 and Table 7. Columns contain expression and their San ≅ 45°
0 9 = MAX(∆x (C))
meaning.
C = Character Ch = Character Height Cw = Character Width

IET Research Journals, pp. 1–11


©The Institution of Engineering and Technology 2019 3

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 7


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

∝ respect to
1 = the angle between( EPe , Bpe )and Z3L baseline
Sw is constant in every row
1 | ∝ −45°| < 𝜎 𝑤ℎ𝑒𝑟𝑒 0
σ = 10°
3
1
∆x (Epoint ) < 𝑀𝐴𝑋(∆x (S)) 1 NWr = 2 End point turn
2
towards right.
3 f Money Minded
1 3 L = 0&𝑁𝑊R T = 1 for v Writing like
Lh > 3 ∗ Lw number, 2
3 2 L = 0&𝑁𝑊R B = 2 for w
1 3 L = 0&𝑂 = 0&𝑁𝑊R T
Sh > 0.6 ∗ Ch &Lw < 0.5 ∗ Lh
4 3 =3
1 3
Sh > 0.6 ∗ Ch &Lw > 0.5 ∗ Lh L = 1&Lh < 0.4 ∗ Ch
5 4
1 3 Irritating to
Ow > Oh NWR T > 3 4 i Slash on top
6 5 others
1
Lh < 0.25 ∗ Ch
7
1 5 p Fork structure Argumentative
H=1
8

End points
Unaggressive in
Table 4. List of hypothesis derived based structural features of 6 q stretches like
profession
handwritten character images for positive social behavior of the character U
person.
Handwritten
NO. Letter Observations Behaviour
Character Use sympathy to
7 s One curbed
cheat

Regular
1 k Person is witty
Character
The bar crosses
Extremely
8 t “t” from top-left
dominating
to bottom-right
Closed loop at
Witty and
2 k left side of the Bar creates Find fault of
Romantic 9 t
stem concavity others

Only one bar for Has extra mental Bar crosses “t” Sarcastic
3 th both the ability to make 10 t
with slash shape Behaviour
character work successful
Crossing by two Not fit for the
11 x
Huge curb at straight lines society
4 y Love with love
lower zone
End point
stretches towards Hide his actual
12 y
up which creates intention
loop
Table 5. Conditions for the hypotheses listed in Table 4.
Loop at lower
NWR R = 3 part and it Worries to reach
1 3 L=1 13 y
stretches towards success
to up.
2 O = 1&𝑁𝑊R R = 4 4 Curbh < 𝑐𝑢𝑟bw
Stem does not
have any turn Sarcastic due
14 y and thickness physical
Table 6. List of hypothesis derived based on contours of deceases weakness
handwritten character images for negative social behavior of the gradually
person.
Retracing from Rigid and takes
Handwritten 15 y
NO. Letter Observations Behaviour end point revenge
Character
Stem of the
Sensitive to Guilty mind
1 d character creates
criticism Opposite stroke (has done
tall loop. 16 y
End point creates Suspicious, at lower zone something
2 E straight line with Sarcasm and wrong)
some angle with irritable

IET Research Journals, pp. 1–11


4 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 8


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

End point creates


character images. This helps us to collect more samples.
Material However, it is hard to verify the response of the user with the
loop with the
17 y glamour to output of the proposed system. At the same time we lose the
stem at lower
drawn attention
zone ground truth of character images.
To ovecome this limitation and for fair evaluation of the
Pretending as
18 z Bar at middle proposed method, we create a second dataset (dataset-2) which
high confidence
includes off-line handwritten character images of 30 writers
with ground truth as the shown sample images and their
Cursive at ground truth in Table 8. In this case, we collect images when
19 z Jealous
beginning the person is present physically. Due to different mechanisms
of image collection, one can guess that dataset-1 contains a
Table 7. Conditions for the hypotheses listed in Table 6 large vartions in image collection, while dataset-2 contains
1 O = 1&Lh > 2 ∗ lw 11 Line = 2&Bpoint = 1
less variations compared to dataset-1. We believe that
evaluating the proposed system on diffeent datasets results in
2 (EPe and Bpe ) makes a straight line 12 NWR T fair evaluations for person behavior identification. More
= 2&&∆y (Epoint ) details about dataset-1 and dataset-2 can be found in Table 1.
< ∆y (Uz ) Since the primary goal of the proposed work is to predict
3 ∆y (Epoint ) < ∆y (S) 13 L = 1&∆y (Epoint ) person behaviors irrespective of age, gender, sex, qualification,
> ∆y (Uz ) paper, pen, ink, application, etc, dataset-1 and dataset-2 do not
4 Epoint 1(Id ) − Epoint 2(Id ) > 0.2 ∗ Ch 14 NWR T = 1 provide the above details.
5 ∆x (O) − ∆x (s) > 0.2 ∗ Cw 15 Run − length of ZL Table 8. Sample images of dataset-2 and their ground
> Sw truth
6 ∆x (Epoint ) > ∆x (Mz ) 16 NWR B = 2 &𝑁𝑊R T
=2
Handwritten Behavior Behavior (user)
image (Graphologist)
7 NWR B = 2 17 L = 1 &𝑁𝑊R B
= 2 &𝑁𝑊R T = 2
Talkative I agree with the
result (YES)
8 ∆x (Bl ) < ∆x (Br ) 18 ∆y (Bl ) = Ch /2
9 ∆y (Bpoint ) > ∆y (Bl )&∆y (Bpoint ) 19 NWR T = 3
> ∆y (Br )
10 Bw (Bl ) > Bw (Br )
I agree with the
Aggressive in result (YES)
profession
4. EXPERIMENTAL RESULTS
To evaluate the performance of the proposed method, we Intelligent and open I agree with the
minded result (YES)
developed an interactive system which allows users to directly
write characters or upload scanned images of handwritten
characters. When a user writes or uploads handwritten
Diplomatic I agree with the
characters, the system automatically displays responses that result (YES)
represent personal behaviors as listed in the Proposed
Methodology Section. Afterwards the user has two options,
either reject prediction as “I don’t agree with the results” or
Analytical Person I agree with the
accept prediction as “I agree with the results”. For each result (YES)
character written by the user, the proposed system predicts
user behaviors. If the user agrees with system decision, we
count it is as one correct. Otherwise, we consider it as a wrong
count. In this case, the performance of the system depends on
user decisions of agree or disagree.
We believe that the user responds by giving his/her
decision to the the proposed system without any bias.
However, sometimes, it is hard to ensure that user response is
genuine for all situations. To overcome this issue, we have
collected a new dataset with clear ground truth for
experimentation. Since this data provides ground truth of
individual person behaviors as the shown sample images and
respective ground truth in Table 8, we can verify the response
given by the user and the proposed system for each instance of
writing. In this work, we consider two datasets, namely,
dataset-1 without ground truth and dataset-2 with ground truth.
Dataset-1 comprises the collection of online handwritten
character images and uploaded images, which are off line Fig. 3. Screen shot of the proposed interactive system to validate
handwritten character images. For online image collection, we the responses against character written by users
send the link of the proposed system to different users to write
IET Research Journals, pp. 1–11
©The Institution of Engineering and Technology 2019 5

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 9


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Fig. 3 shows sample screen shots of the proposed The accuracies of the proposed features + rules, CNN, SVM and
interactive system, where we can see “upload” option to get existing methods are reported in Table 10, where it is noted that
handwritten characters by scanning, “clear” option to remove the proposed features + rules achieve the best accuracy
the previous results displayed on screen, “analyze” option to compared to all the other methods. It is observed from Table 10
send request to the system to test characters, “I don’t agree that the proposed and existing methods score better results for
with the result” option is to reject the response given by the dataset-2 compared to dataset-1. This is due to more samples
system for the written characters, and “I agree with the results” with large variations in image collection compared to dataset-2.
option is to accept the responses given by the system for the It is also true that writing on paper using pens gives more natural
characters. Sample response for character “a” written by the writing of individuals than writing using online devices. When
user is displayed as “the respective person is talkative”. This we compare the proposed features + rules with the proposed
is one sample of the particular person. features + CNN and the proposed features + SVM, the proposed
features + rules are better than the other two methods for both
To measure the performance of the proposed method, we
dataset-1 and dataset-2 as reported in Table 10. The reason is
use accuracy, which is defiend as the total number of correct
that the proposed method derives rules for each person behavior
responses given by user (agree or disagree for the proposed
identification according to graphologist, and the rule are unique
system prediction) divided by the total number of characters.
in nature and do not overlap with other rules. Moreover, the
To show usefulness of the proposed method, we implement the
proposed shape-based features are invariant to different
state-of-the-art methods, namely, the method [11] which
variations caused by pen, paper, ink, device, online writing, off-
defines rules as the proposed method for human behavior
line writing, etc. It is evident from the results shown in Fig. 4(a),
prediction, and one more method [12] which explores an
where it is seen that as long as the structure of a character is
artficial neural network for human behavior prediction, for
preserved, the proposed rules work well.
comparative study with the proposed method. The main reason
to use these two methods for comparative study is that these Table 10 shows that the proposed features with SVM is
two are the state of the art methods and set the same objecive better than the proposed features with CNN. This is valid
as the proposed method. In the same way, to assess the because of free style writing as one can expect large variations
contribution of the feature extraction and defined rules, we in writing. Besides, since users use special pen and pad for
pass the extracted features to CNN and SVM classifier for writing characters, which is not usual practice of writing, we can
person behavior identification. The number of samples for expect still more variations than writing on paper. When we
training and testing for the dataset-1 and dataset-2 are listed in have large variations, CNN requires more samples to achieve
Table 9. The same set up is used for all the experiments in this better results as it helps to determine proper weight for
work. For SVM classifier, we follow the instructions in [16] classification. However, collecting a large number of samples
and for CNN, we use the pre-defined architecture as in [17] for for this work is hard because people do not like to share their
experimentaion. personal behaviors. On the other hand, SVM does not require
more samples for classification if a feature extracts the unique
For implementing CNN, we use the architecture proposed
property of character images of different writings. Therefore, for
in the method [17] for image recognition, which is called VGG
our datasets, the proposed features with SVM gives good results
architecture. The details of the architecture are as follows.
compared to those of features with CNN.
VGG architecture involves layers in the first, second, third,
fourth and fifth convolution blocks containing 64, 128, 256, It is also noted from Table 10 that the existing methods report
512 filters, respectively. All the layers use filter size of 3X3 poor results for both the datasets compared to the proposed method
with ReLU activation. After the fifth convolution block, we including the classifier based methods. The main reason is that the
flatten the output and add three dense layers with 1024,1024 existing methods focus on particular applications and specific
and 59 nodes, respectively. The first and second dense layers character shapes for person behavior identification. As a result, the
have ReLU activation, whereas the third dense layer has existing methods may not cope with the complexity of the proposed
Softmax activation. We use 0.5 dropout between two dense problem. The method in [12] extracts rule-based features for the
layers to reduce chances of overfitting. The model contains specific character “y” to identify person behaviors. For fair
comparative studies, the proposed rule based method is tested on
16M trainable parameter. We train the network using SGD
the same character “y” to calculate accuracy. The results are
optimizer with learning rate 0.0001, momentum 0.5 and batch
reported in Table 10 for both the datasets. It is observed from Table
size 50. More details can be found in [17]. 10 that the proposed rule based method is better than the existing
Table 9. Statistical details of the dataset for evaluation method [12] for both the datasets. The reason for the poor results of
the existing method [12] is that the features extracted are not robust
Datast type Total Number of Number of Total number as the proposed features.
number training testing samples of samples
of writers samples It is true that most of the time, persons use lower case letters
for handwriting and seldom use capital letter for writing. As a result,
Dataset-1 600 3700 900 4600 most of the characters in the dataset-1 and dataset-2 are lower case
Dataset-2 30 500 200 700 letters. However, sometimes, users may write capital letters. To test
the proposed method on capital letters, we pass both lower and
upper case letters to the system as shown in Fig. 4(a), where it can
4.1 Evaluating the Proposed Person Behavior be seen that the proposed method predicts the same behavior for
Identification Method both lower and upper case letters, “W”. Since the proposed system
IET Research Journals, pp. 1–11
6 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 10


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

(a) The proposed system predicts the same behavior for both upper case and lower case letters

(b) The proposed system predicts different behavior for upper case and lower-case letters.
Fig. 4. The effectivenes of the proposed system for upper case and lower caset letters.

extracts features based on character shapes to predict person Dataset- 69.95 61.60 67.25 51.0 99.95 59.0
behaviors, as long as the structure of a character is preserved, the 1
proposed system works well as shown in Fig. 4(a). If shapes and Dataset- 86.70 69.55 78.75 53.0 92.70 62.0
2
structures of upper and lower case letters are different, the proposed
method predicts different behaviors as shown in Fig. 4(b), where
one can see different behaviors for upper and lower cases of the
same character. This is true because when shape and structure
change, the rule for person behavior also changes according to
graphology.

0.0 0.2 0.4 Fig. 6. The performance of the proposed system for noisy
images

Similarly, the proposed method is also tested on noisy


images to assess performance. Sample images with added
0.6 0.8 0.9 Gaussian noises manually at different levels are shown in Fig.
Fig. 5. Sample images affected by Gaussian noise at different 5, where one can see as Gaussian noise level increases, noise
levels. density (the number of Gaussian noise pixels) increases. For
images with Gaussian noises, we calculate accuracy using
Table 10. Accuracy of the proposed and existing methods proposed feature with rules, the proposed features with SVM,
for both the datasets (in %) and the proposed features with CNN as shown in Fig. 6. It is
noted from Fig. 6 that the proposed features with rules, SVM
Datasets Proposed + Proposed + Proposed + [11] Specific characters and CNN do not work well for noise images as we can see the
Rule CNN SVM “y”
Proposed + [12] performance decrease as noise level increases. Therefore, one
Rule can argue that the proposed feature are not robust to noise.
IET Research Journals, pp. 1–11
©The Institution of Engineering and Technology 2019 7

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 11


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Note that since we collect images online using pen and pad, ratio, height, width and shape of written characters. According
the process does not introduce any noises. As a result, the to graphology, the proposed method derives hypothesis to
above limitation cannot be considered as a drawback of the identify personal behaviors. In this work, we consider possible
proposed method. personal behaviors for identification. To validate the proposed
method, we have developed an interactive automatic system to
Though we propose an effective system for person
test the hypothesis, which accepts written characters from
behavior idenfication, sometimes, it fails to predict actual
different persons to predict specific behaviors. The system also
person behaviors based on handwriting analysis. This is due to
allows users to choose accept or reject for each predicted
subjectivity of individual persons. In other words, it is hard to
behavior. Experimental results and a comparative study on two
predict exact behaviors using any mode becase there is no
datasets show that the proposed method outperforms the
boundary for defining person behavoirs. In this case, the
existing methods in terms of accuracy. Furthermore, we can
proposed system predicts incorrect behaviors of a person. So,
also conclude that the proposed features + rules achieve the
a user selects “I don’t agree with the results” as shown in Fig.
best accuracy compared to the proposed features + CNN and
7(a), where the predicted behavior does not match with the
the proposed features + SVM.
user. Sometimes, if the written character may not be in defined
repository, there are chances of losing accruacy. In this case, When a person writes touching characters, the
the proposed system suggests the user to choose another performance of the proposed system degrades because it
character or rewrite the same character again as shown in Fig. accepts individual characters as the input for person behavior
7(b), where written character features do not match with the identification in this work. Therefore, we have planned to
corresponding character. combine character segmentation from handwritten text line
and person behavior identification in the future. Sometimes,
when a person writes characters with more variations in shape,
the extracted features may overlap with features of other
characters. This leads to poor results for the proposed system.
Therefore, the proposed system requires high level features,
such as context between successive characters in words to
improve the results. It is true that the rules defined based on
graphologist are limited to the number of person behaviors. In
order to improve the performance of the proposed method for
overcoming the above-mentioned limitations, we plan to
explore different CNN architectures using online time
information of the writings in the near future.
ACKNOWLEDGMENTS
(a) Example for “I don’t agree with the results” The work described in this paper was supported by the
Science Foundation for Distinguished Young Scholars of
Jiangsu under Grant No. BK20160021, the Natural Science
Foundation of China under Grant No. 61672273 and No.
61272218. The authors acknowledge the help of Mr. Aishik
Konwer and Mr. Abir Bhowmick of Department of ECE,
Institute of Engineering & Management, Kolkata in this work.

REFERENCES
[1] R. P. Tett and C. A. Palmer, “The validity of handwriting elements in
relation to self-report personality trait measures”, Person Individual
Difference, Vol. 22, pp 11-18, (1997).
[2] J. F. Brewer, “Graphology”, Complementary Therapies in Nursing &
Midwife, 5, pp 6-14, (1999).
[3] L. G. Hafemann, R. Sabourin and L. S. Oliveira, “Analyzing features
learned for offline signature verification using deep CNNs”, In Proc.
(b) Example for not finding any match with the ICPR, pp 2989-2994, (2016).
defined behaviours of the person. [4] O. M. Hurtado, R. Guest, S. V. Stevenage and G. J. Neil, “The
relationship between handwritten signature production and
Fig. 7. Sample incorrect results for lowering accuracy. personality traits”, In Proc. IJCB, pp 1-8, (2014).
[5] V. Gattupalli, P. S. Chandakkar and B. Li, “A computational approach
to relative aesthetics”, In Proc. ICPR, pp 2446-2451, (2016).
[6] A. Majumdar, P. Krishnan and C. V. Jawahar, “Visual aesthetic
5 CONCLUSIONS AND FUTURE WORK analysis for handwritten document images”, In Proc. ICFHR, pp 423-
428, (2016).
We have proposed an automatic system for identifying [7] S. V. Kedar and D. S. Bormane, “Automatic personality assessment:
human behaviors based on handwriting at character level by A systematic review”, In Proc. ICIP, pp 326-331, (2015).
keeping real time applications of graphology. The proposed [8] B. Fallah and H. Khotanlou, “Identify human personality parameter
method extracts structural features such as slant, hole, aspect based on handwriting using neural network”, In Proc. IRNOPEN, pp
120-126, (2016).
IET Research Journals, pp. 1–11
8 ©The Institution of Engineering and Technology 2019

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 12


Auto-generated PDF by ReView
This article has been accepted for publication in a future issue of this CAAI
journal,Transactions onfully
but has not been Intelligence
edited. Technology
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

[9] S. Asra and D. C. Shubhangi, “Human behavior recognition based on


handwritten cursives by SVM classifier”, In Proc. ICEECCOT, pp
260-268, (2017).
[10] M. Topaloglu and S. Ekmekci, “Gender detection and identifying
one’s handwriting with handwriting analysis”, ESWA, Vol. 79, pp
236-243, (2017).
[11] H. N. Champa and K. R. A. Kumar, “Artificial neural network for
human behavior prediction through handwriting analysis”, IJCA, Vol.
2, pp 36-41, (2010).
[12] H. N. Champa and K. R. A. Kumar, “Automated human behavior
prediction through handwriting analysis”, In Proc. ICIIC, pp 160-165,(
2010).
[13] R. Coll, A. Frnes and J. Llados, “Graphological analysis of
handwritten text documents for human resource recruitment”, In Proc.
ICDAR, pp 1081-1085, (2009).
[14] U. Pal, A. Belaid and Ch. Choisy, "Touching numeral segmentation
using water reservoir concept", Pattern Recognition Letters, vol.24,
pp. 261-272, (2003).
[15] Marc Seifer, “The Complete Book of Handwriting Analysis”,
Pentagon Press, (2014).

[16] B. E. Boser, I. M, Guyon and V, N, Vapnik, V.N, “A Training


Algorithm for Optimal Margin Classifiers”, In Proc. COLT, pp 144-
152, (1992).
[17] K. Simonyan and A. Zisserman, “Very deep convolutional networks
for large-scale image recognition”, In Proc. ICLR, pp 1-14, (2015).

IET Research Journals, pp. 1–11


©The Institution of Engineering and Technology 2019 9

CIT-2019-0051-Main-R2-2-UP.pdf M ainDocument IET Review Copy Only 13

You might also like