Fuzzy rule based classification and recognition of handwritten hindi

INTERNATIONALComputer EngineeringCOMPUTER ENGINEERING
International Journal of JOURNAL OF and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME
& TECHNOLOGY (IJCET)
ISSN 0976 – 6367(Print)
ISSN 0976 – 6375(Online)
Volume 4, Issue 1, January- February (2013), pp. 337-357
IJCET
© IAEME:www.iaeme.com/ijcet.asp
Journal Impact Factor (2012): 3.9580 (Calculated by GISI) ©IAEME
www.jifactor.com

FUZZY RULE BASED CLASSIFICATION AND RECOGNITION
OF HANDWRITTEN HINDI CURVE SCRIPT

Gunjan Singh1, Avinash Pokhriyal1, Sushma Lehri2
1
( Faculty of Management & Computer Application, RBS College, Agra, India.)
2
(Professor, I ET, Dr. B. R. Ambedkar University, Agra, India.)

ABSTRACT

This paper presents a novel system for classification and recognition of
handwritten Hindi script using fuzzy rule based approach. Classification & recognition of
handwritten Hindi script is a complex task as characters are cursive in nature and
demonstrate a lot of similar features. The quality of fuzzy logic to deal with vague and
imprecise data makes it appropriate for such problems. In this paper, we focus on two or
three letter words without modifiers. Prior to recognition, handwritten words are
preprocessed and segmented into individual characters. The performance of an optical
character recognition system extremely depends on the procedure used to extract quality
features from characters. During classification stage characters are classified into seven
classes using fuzzy if-then rules based on one of the most important component of Hindi
characters – the vertical bar. Features such as curves, lines, junction points and endpoints
are used at the recognition stage. A 3x3 mask is used to extract features from character
image. System was tested for total 450 words written by 30 different people.
Experimental results show that the proposed method performs classification and
recognition at the rate of 92.02%. The proposed system has been implemented in
MATLAB 2009 environment.

Keywords: Classification, Fuzzy rule based approach, Handwritten Hindi curve script,
Vertical bar, 8-neighbourhood

337

International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-

1. INTRODUCTION

Character recognition is a broad field in which all types of machine recognition of
characters in various application domains is studied. It includes the recognition of machine
printed as well as hand written characters. Recognition of machine printed characters
involves the recognition of characters written by a machine, while handwritten character
recognition includes the recognition of characters written by human being either online or
offline. Recognition of machine printed characters is easy as characters are of same size, font
& thickness and have a proper shape, but due to various writing styles, hand written character
recognition is difficult as characters may be of different sizes, width and orientation. A
comparison of both approaches is given in [1]. In this paper, we will present a fuzzy rule
based classification and recognition system for handwritten Hindi script.

Hindi is one of the official languages of India. It is world’s third most commonly used
language after Chinese and English. Hindi script has 13 vowels (‘SWARS’) and 33
consonants (‘VYANJANS’) in its basic character set. All the characters have two common
features – (i) their cursive nature and, (ii) presence of header line (‘SHIROREKHA’). Header
line is a powerful tool of Hindi language. These features differentiate
the script from English and other Latin scripts. Words are formed by combining characters,
half characters and /or modifiers using header line. Fig.1 shows basic character set, a list of
modifiers and few words.

(b)

(a) (c)

Figure 1(a). Basic character set, (b) Swars (vowels) & corresponding matras
(modifiers) and (c) Few Hindi language words

Now-a-days Hindi is being used worldwide in many fields such as banking, medical,
science and technology etc. Most of the Hindi language words are being included in world’s
best dictionaries and other vocabulary developing tools. Due to the increasing popularity,
automatic Hindi language recognition systems have now become important. Research in this
area started in early 1970s. In 1977, Sethi and Chatterjee [2] presented a constrained
recognition system for handwritten Hindi characters. In [3], Sinha and Mahabala presented a
syntactic pattern analysis system for the recognition of machine printed and handwritten
characters. The first complete OCR system for machine printed characters is presented in [4].
Recognition of handwritten Hindi characters is still difficult for a machine as characters are

338


cursive in nature and show a lot of similarities such as presence of header line, presence /
absence of vertical bar, loops & curves. A survey for handwritten character recognition was
proposed by R. Srihari [5] in 2000. Most of the work is focused on the recognition of
individual characters, and a little attention has been paid towards the recognition of words,
sentences or text. Recognition of words is difficult as words should be segmented into
individual characters. In the present paper, we propose a fuzzy rule based classification and
recognition system for handwritten Hindi curve script words of two or three letters without
modifiers.

Fuzzy logic is an organized method to solve problems dealing with vague, ambiguous,
imprecise, noisy, or missing input data. The concept of fuzzy logic is first given by Dr. Lotfi
A. Zadeh in 1965[13]. According to Dr, Zadeh, fuzzy logic is a mathematical tool for dealing
with uncertainty. As compared to crisp logic that deals with precise values; it is a form of
multi valued logic, which provides a way to deal with reasoning that is approximate. So it
gives a machine a better mean to simulate human reasoning capabilities. Dealing with
approximation makes it appropriate for problems such as handwritten character recognition.
This paper is organized in 5 sections. Section 2 throws some light on work done in the field
of handwritten Hindi character recognition. Section 3 presents the proposed system. Section 4
shows the experimental results. Finally conclusion is made in the last section.

2. LITERATURE REVIEW

Hanmandlu et al. [6] presented a fuzzy model based recognition system for
handwritten Hindi characters with 90.65% accuracy. The system works by performing coarse
classification of preprocessed character image by dividing it into 3x3 windows and then
determining the presence and position of vertical bar. Then feature are extracted by applying
the box approach. For recognition, an exponential variant of fuzzy membership function,
constructed using the normalized vector distance, is used. Mukherjee and Rege [7] presented
a shape feature and fuzzy logic based offline handwritten character recognition system for the
language with 86.4% recognition rate. Structural features, such as end points, junction points,
and adaptive thinning algorithm are used for segmenting characters into strokes. Then crisp
and fuzzy features are extracted for each stroke of the character. Two stage classification is
performed. Pre classification is performed using tree classifier in which characters are
classified based upon the presence and position of vertical line. Final classification and
recognition is performed using unordered stroke classification based on mean stroke features.
In [8], a handwritten Hindi vowel character recognition system is presented, in which vowels
are segmented into five groups using projection approach. To extract the core character
header line is removed by applying horizontal projection and modifiers are removed using
vertical projection. Feature extraction is done by using Invariant moments. Holambe and
Thool [9] presented a system for the recognition of printed and handwritten Devanagari script
using support vector machine and k-nearest neighbour classification technique. Singh, Mittal
and Ghosh [10] perform estimation of Support vector machine with Radial basis function and
k-nearest neighbour and achieved 93.8% accuracy. Two methods – curvelet transform &
character geometry used for extracting features.

339


3. PROPOSED SYSTEM

The proposed system works in six stages: preprocessing, segmentation, normalization,
classification, feature extraction and recognition. Flow diagram is shown in Fig.2.

Preprocessing
Start

Slant
Thinning Binarization Correction Dilation Erosion Filtering Scanning

Noise Reduction

Feature
Segmentation Normalization Classification Extraction Recognition

Figure 2. Flow diagram of the proposed

3.1 Preprocessing
During preprocessing, a number of following operations are performed on the collected data
to make it suitable for further processing—
(i) Scanning— Handwritten word data samples, collected from various people, are
scanned through an optical scanner or camera to convert data into a gray scale
image.
(ii) Noise Reduction-- Noise may be introduced in image during scanning, so to
reduce noise following operations are performed:
(a) Filtering—to reduce noise and false points, a nonlinear spatial filter- median
filter is applied. Concept is to convolute a predefined mask with the image and
replaces the value of the centre pixel by the median of intensity values in the
neighbourhood of that pixel [14]
(b) Dilation— there may be gaps in characters, which are filled by dilation using a
structuring element [14].
(c) Erosion— to eliminate the spurious objects from the image, erosion is applied
on it.
(iii) Slant Correction— there are chances that characters in the word are inclined
upwards or declined downwards, which makes feature extraction process difficult.
For that, slant correction is done by using [ 12].
(iv) Binarization--In this paper, features are extracted from binary images of
characters, so there is a need to convert the image to binary form. Global
thresholding is applied for binarization. The method works by choosing a
threshold value for the whole image and then sets the values of pixels to 1 whose
value is greater than the threshold and 0 otherwise.
(v) Thinning—Finally, binary image is thinned to single pixel width by the method
presented in [11].

340


3.2 Segmentation
Thinned image of word is segmented into individual characters by histogram equalization as
following—
(i) First, horizontal histogram is taken to get the upper and lower boundary of the
word.
(ii) Then vertical histogram is taken to get the region of each character.
(iii) A case occurs when number of regions is more than the number of characters in
the word. It may be due to the presence of a character in which vertical bar is not
connected to the character. In that case, the region of the vertical bar, with highest
peak value, is considered to be a part of the character to its left.

3.3 Normalization
Binary images of individual characters are normalized into 9x9.

3.4 Classification
All Hindi language characters are made up of mainly three components: header line or
SHIROREKHA, vertical bar, and curves. In the proposed method, we choose vertical bar
component to classify characters. TABLE 1 shows the features (presence or absence, length,
position, connectedness of vertical bar and number of junction points) on which basis
different classes of characters are formed. A character can belong to one class only.

Table 1: Features used for classification
Feature Symbol Values
P (present)
Presence of vertical bar VB
NP(not present)
M(middle)
Position of vertical bar POS
RE (right end)
S (20%-30% of the character width W)
Length of vertical bar LEN
L(70%-80% of the character width W)

Connectedness of vertical bar to C (connected)
character CON
NC (not connected)

Number of junction points JP 1,2,3.4, or 5

A junction point is a point with 3 or more pixels in its neighbourhood .Method of
extracting these features is given in algorithm VERTICALBAR_INFO and
JUNCTIONPOINT_COUNT. A movable 3X3 mask (Fig.3) is applied on the image, which
shows 8-neighborhood of the pixel P0.

341


P8 P1 P2
P7 P0 P3
P6 P5 P4
Figure 3: 3X3 mask
In these algorithms, following notations are used:
CP -- current pixel
CL -- current location
COUNT_1 -- counter variable to count the number of pixels. Initial value is set to 0.
COUNT_2 -- counter variable to count the number of junction points. Initial value is set to 0.
ROW -- current row number
COL -- current column number

Algorithm VERTICALBAR_INFO

To determine the information about the vertical bar do the following:
1. Starting from the last column of the first row i.e. ROW==0 & COL==8, convolute
the mask on the binary image of character and check:
(i) IF pixel is a foreground pixel then call it as P0.
IF number of neighbouring pixels of P0 ≥ 3 and one pixel is P5 then do
the following --
(a) Set CP = P0.
(b) Set N = COL.
(c) Increase COUNT_1 by 1.
(ii) ELSE move to next column to the left and repeat step (i) till COL ≥ 4

2. To identify the presence of vertical bar check the value of COUNT_1
IF COUNT_1 ==1
THEN VB is P
ELSE VB is NP.
3. To identify the position of vertical bar check the value of N.
IF N ≥ 8
THEN POS is RE
ELSE POS is M
4. To identify the length and connectedness of vertical bar to character check POS.
(i) IF POS==M
THEN do the following till P5 is encountered
(a) Set P5=P0
(b) Increase COUNT_1 by 1
(ii) IF COUNT_1 >3
THEN LEN is L
ELSE LEN is S
(iii) IF POS ==RE
THEN Set CP=P0 and check the following till P5 is encountered
IF P6 OR P7 OR P8 exists
THEN CON is C
ELSE CON is NC

342


Algorithm NUM_JUNCTIONPOINTS

To determine the number of junction points do the following

1. Starting from the upper left corner pixel, convoluting the mask on the
image from left to right.
2. Find the first foreground pixel P0
IF number of neighbouring pixels of P0 ≥ 3
THEN increase COUNT_2 by 1
ELSE P0=P3
3. Repeat step 2 till rightmost lower pixel is obtained.
4. Set JP=COUNT_2

Using above mentioned algorithms, following fuzzy rules are formed to classify the
characters into one of the eight classes. Flow process is shown in Fig.4.

(i) IF VB == NP THEN character belongs to class A ( )

(ii) IF VB == P AND POS == M AND LEN == L THEN character belongs to class B
( )

(iii) IF VB == P AND POS == M AND LEN == S AND JP < 2 THEN character
belongs to class C( )

(iv) IF VB == P AND POS == M AND LEN == S AND JP ≥ 2THEN character belongs
to class D ( )

(v) IF VB == P AND POS== RE AND CON == NC THEN character belongs to class
E ( )

(vi) IF VB == P AND POS == RE AND CON == C AND JP <4 THEN character
belongs to class F( )

(vii) IF VB == P AND POS == RE AND CON == C AND JP ≥ 4 THEN character
belongs to class G( )

343


Read normalized image of size 9X9 VB : Vertical bar
of the character A: Absent
POS : Position of vertical bar
RE : Right end
Read presence of VB
M: Middle
Character
LEN : Length of vertical bar
belongs to class yes L : Large
A If VB==A S: Small
( ) JP : Junction point
no CON: Connectedness of vertical bar
Read position of VB NC : Not connected

yes
If POS==RE Read connectedness of VB

no
Read length of VB yes Character
If belongs to
CON==NC class E
Character yes ( )
belongs to class If LEN==L
B
( ) no
no

Read value of JP Read value of JP

Character belongs yes
If If JP ≥ 4
to class
D JP ≥2
( ) no yes
no
Character belongs to Character belongs to
Character belongs to class C class F class G
( ) ( ) ( )

Figure 4. Flow process of classification

3.5 Feature Extraction
Steps for extracting features are given in following algorithm--
Algorithm FEATURE_REC
1. Remove header line by applying the following method-
(i) Apply the 3X3 movable mask on the normalized image and scan the first row
from right to left.
(ii) IF pixel is a foreground pixel then call it P0.
IF P7 is a foreground pixel OR P0 is an end point OR P0 is a
disconnected component
SET P0 = 0
ELSE move to the left pixel.

344


Image is scanned from right to left to avoid the deletion of character pixels in
characters such as: because these characters, except , may be written
in two ways— (a) header line covers the whole character and, (b) when header line
covers only half or a portion of the character. In the first case, this step may result in
deletion of pixels, which are common to header line and character, in characters
mentioned above as well as characters such as and may produce some
disconnected components with small number of pixels.

2. Delete disconnected components as following--
(i) Scan the second row of the image from left to right.
(ii) Find the first foreground pixel P0.
(iii) IF P3 ==1
IF any pixel in 8 neighbourhood of P3 does not exists
THEN SET P0=0 AND P3=0
ELSE IF P5==1
IF any pixel in 8 neighbourhood of P5 does not exists
THEN SET P0=0 AND P5=0

Fig. 5 shows the process of deleting header line from character and its result.

(a) (b) (c)

Figure 5: (a) Character with header line, (b) Character without header line and
disconnected component, (c) Character after removing disconnected component

3. Apply the 3X3 movable mask on the normalized image of classified character and
scan the image
from top to bottom row wise. Collect following information for junction points and
end points--
(i) N1 : total number of junction points
(ii) N2: total number of end points
(iii) JPi : ith junction point, where i=1 to N1
(iv) EPi : ith end point where i=1 to N2
(v) Curve (JPi) : curve on ith junction point (Table 2)
(vi) Curve (EPi) : curve on ith end point
(vii) Line(JPi) : line on ith junction point (Table 2)

345


(viii) Line(EPi) : line on ith end point
(ix) Loop(JPi) : loop on ith junction point
(x) D1(i): direction of next endpoint from ith end point
(xi) D2(i): direction of next junction point from ith junction point

Values and symbols of different types of curves, lines & loops are given in the
TABLE 2.

Table 2: Values and symbols for curves, lines and loop
Features Values Symbol

Left Curve LC

Upper left curve ULC

Lower left curve LLC

Curve Right curve RC

Upper right curve URC

Lower right curve LRC

U curve U

Vertical line VL

Horizontal line HL
Line
Back slash BS

Present P
Loop

Not present NP

Different forms of above mentioned curves, lines and loops are shown in Fig. 6.
In this code, following notations are used:

PS -- Starting point
CL -- current location
CP -- Current pixel
COUNT -- counter variable. Initial value is set to 0.

346


Algorithm CURVE_LINE_LOOP_INFO

To determine the nature of the curve do the following:
Convolute the mask on the binary image of classified character from bottom to top row wise.
Let P is the first foreground pixel. Call it current pixel (CP).
1. If CP is a junction point or end point, then check the 8-neighbourhood of CP.

(a) IF P1 is true THEN
(i) Repeat till P1 is encountered
(ii) Increase COUNT by 1.
ELSE stop.
(b) IF P3 is true THEN
ELSE stop.
(c) IF P8 is true THEN
ELSE stop.
(d) IF P1 OR P2 is true THEN
(i) Repeat till P1 OR P2 is encountered
ELSE stop.
(e) IF P1 OR P8 is true THEN
ELSE stop.
(f) IF P2 OR P3 OR P4 is true THEN
(i) Repeat till P2 OR P3 OR P4 encountered
ELSE stop.
(g) IF P4 OR P5 is true THEN
ELSE stop.
(h) IF P6 OR P7 OR P8 is true THEN
(i) Repeat till P6 OR P7 OR P8 is encountered
ELSE stop.

2. Check the following to know the type of curve and line:

(i) IF step 1(h) is true
IF step 1(a) is true
IF step 1(f) is true
IF COUNT ≥ 3
THEN Curve is LC

347


(ii) ELSEIF step 1(e) is true
IF COUNT ≥2
THEN Curve is ULC
(iii) ELSEIF step1(h) is true
IF step 1(e) is true
IF COUNT ≥2
THEN Curve is LLC
(iv) ELSE IF step 1(f) is true
IF step 1(a) is true
IF step 1 (h) is true
IF COUNT ≥ 3
THEN Curve is RC.
(v) ELSE IF step 1(d) is true
IF step 1(h) is true
IF COUNT ≥ 2
THEN Curve is URC.
(vi) ELSE IF step 1 (f) is true
IF step 1(e) is true
IF COUNT ≥ 2
THEN Curve is LRC.
(vii) ELSEIF step 1(g) is true
IF step 1(h) OR step1 (f) is true
IF step 1(d) is true
IF COUNT ≥3
THEN Curve is U
(viii) IF step 1(a) is true
IF COUNT ≥ 2
THEN Line is VL
(ix) IF step 1(b) is true
IF COUNT ≥ 2
THEN Line is HL

(x) IF step 1(c) is true
IF COUNT ≥ 2
THEN Line is BS

3. If CP is a junction point, then do the following to check the presence of loop:
IF step 1(h) is true
IF step 1(a) OR step 1 (g) is true
IF Pi == CP AND COUNT ≥ 5
THEN Loop is P.

348


CP

(a) (b)

(c) (d)

(e) (f)

(g)

CP

(h) (i)
Figure 6 : Different types of curves : (a) Left curve (LC), (b) Upper left curve (ULC) , (c) Lower
left curve (LLC), (d) Right curve (RC), (e) Upper right curve (URC), (f) Lower right curve (LRC), (g)
U curve (U) , (h) Vertical line (VL), Horizontal line (HL), Backward slash (BS), (i) loop

349


3.6 Recognition
Fuzzy rules are used for recognition. Class wise rules applied for characters are:

1. IF Class is A
IF Curve (EP1) == RC
THEN character is
ELSE IF Curve (JP1) ==LRC
IF N2==4 OR D1 (3) == P3
THEN character is
ELSE character is

2. IF Class is B
IF Curve (EP2) == LC
THEN character is
ELSE IF Curve (EP2) == URC
IF Curve (JP1) == LC OR Loop(JP1) ==P
THEN character is
ELSE character is

3. IF Class is C
THEN character is
ELSE IF Curve (EP1) == RC
IF N2==3
THEN character is
ELSE character is

4. IF Class is D
THEN character is
ELSE IF Curve (JP1) == LC
IF N2 < 2
THEN character is
ELSE IF N2==2
THEN character is
ELSE character is
ELSE IF Loop (JP1) ==P
IF Curve (JP1) == RC OR URC
THEN character is
ELSE character is

5. IF Class is E
IF Loop (JP1) ==P
IF N1==2
THEN character is
ELSE character is
ELSE IF Curve (EP1) == U
THEN character is

350


6. IF Class is F
IF N2 > 3
IF Curve (EP1) == ULC
THEN character is
ELSE IF Curve (EP1) == RC OR Curve (EP2) == RC
THEN character is
ELSE character is
ELSE IF N2==3
IF Curve (JP1) == LLC
THEN character is
ELSE IF Curve (JP1) ==U
THEN character is
ELSE IF Curve (EP1) == ULC
THEN character is
ELSE character is
ELSE
IF Curve (JP1) ==U
THEN character is
ELSE IF Curve (JP1) ==LLC
THEN character is
ELSE IF Curve (JP1) ==LC OR Loop (JP1) ==P
THEN character is
ELSE character is

7. IF Class is G
IF N2>4
IF Curve (EP1) == RC
THEN character is
ELSE IF Line (EP1) == BS
IF D2 (1) ==P3 OR D2(2)==P3
THEN character is
ELSE character is
ELSE IF N2 ==4
IF Loop on JP1 ==P
THEN character is
ELSE character is
ELSE
IF Curve (JP1) ==LLC OR U
THEN character is
ELSE IF Curve (JP1) == LC
THEN character is
ELSE IF Loop on JP1 ==P
IF Loop on JP3 ==P OR LINE (EP2) == HL
THEN character is
ELSE character is

351


Table 3: Summary of fuzzy rules for each character
Class N1 N2 Curve(JP) Curve(EP) Line(JP) Line(EP) Loop(JP) D1 D2 D3 Character
--- -- --- RC --- --- --- --- --- ---
A --- --- --- --- LRC --- --- --- --- ---
--- 4 --- --- LRC --- --- P3 --- ---
--- --- --- LC --- --- --- --- --- ---
B --- --- --- URC --- --- --- --- --- ---
--- --- LC URC --- --- P --- --- ---
--- --- --- LC --- --- --- --- --- ---
C --- --- --- RC --- --- --- --- --- ---
--- 3 --- RC --- --- --- --- --- ---
--- --- --- LC --- --- --- --- --- ---
--- <2 LC --- --- --- --- --- --- ---
--- 2 LC --- --- --- --- --- --- ---
D
--- --- LC --- --- --- --- --- --- ---
--- --- --- --- --- --- P --- --- ---
--- --- RC OR --- --- --- P --- --- ---
URC
2 --- --- --- --- --- P --- --- ---
E --- --- --- --- --- --- P --- --- ---
--- --- --- U --- --- --- --- --- ---
--- >3 --- --- --- --- --- --- --- ---

--- >3 --- ULC --- --- --- --- --- ---
--- >3 --- RC --- --- --- --- --- ---
--- 3 --- --- --- --- --- --- --- ---
F
--- 3 LLC --- --- --- --- --- --- ---
--- 3 U --- --- --- --- --- --- ---
--- 3 --- ULC --- --- --- --- --- ---
--- <3 --- --- --- --- --- --- --- ---
--- <3 U --- --- --- --- --- --- ---
--- <3 LLC --- --- --- --- --- --- ---
--- <3 LC --- --- --- P --- --- ---
--- >4 --- --- --- --- --- --- --- ---
--- >4 --- RC --- --- --- --- --- ---
--- >4 --- --- --- BS --- P3 P3 ---
--- 4 --- --- --- --- --- --- --- ---
G
--- 4 --- --- --- --- P --- --- ---
--- <4 LLC OR --- --- --- --- --- --- ---
U
--- <4 LC --- --- --- --- --- --- ---
--- <4 --- --- --- HL P --- --- ---
--- <4 --- --- --- --- P --- --- ---

352


4. EXPERIMENTAL RESULTS

Dataset has been created by collecting handwritten word samples by 30 people of
different age groups. Each person was asked to write 15 predecided words. A part of dataset
is shown in the following figure—

Figure 7: Word samples taken for experiment

353


These word samples were scanned, using a flat-bed scanner at 300dpi. Results of operations
performed during recognition process on scanned image of word are shown in the
following figure.

Original image

Filtered image

Eroded and dilated image

Binarized image

Thinned image

Segmented
image

VB == P VB == NP VB == P
POS == RE POS == RE
CON == C CON == C
Classification
JP ≥ 4 JP < 4

Character belongs to class G A F

Figure 8. Result of operations performed during preprocessing, segmentation
and classification on sample word

354


After classification, features mentioned in TABLE 2 are extracted for each character by
applying algorithm FEATURE_REC, which are then used at the time of recognition.
Recognition rate for each word sample and for the proposed method is given in TABLE 4.

Table 4. Average recognition rate of selected words
Sample Word Recognition Recognition Recognition Avg.
rate rate of rate of recognition
of character character 2 character 3 rate
1
S1 92.15% 94.08% 88.23% 91.48%

S2 94% 90.11% 87.23% 90.44%

S3 90.93% 97.26% 95.06% 94.41%

S4 94.14% 90.17% 90% 91.43%

S5 83.66% 93.96% 92.07% 89.89%

S6 95% 93.48% 84.36% 90.94%

S7 95.22% 92.01% 89.76% 92.33%

S8 96.31% 92.45% 91.19% 93.31%

S9 88.42% 92.31% 94.21% 91.64%

S10 89.75% 83.52% 93.46% 88.91%

S11 90.68% 88.99% ---------- 89.83%

S12 96.29% 94.43% --------- 95.36%

S13 88.57% 93.91% --------- 91.24%

S14 96.81% 97.44% -------- 97.12%

S15 87.41% 96.80% -------- 92.10%

Overall Average Recognition Rate 92.02%

355

International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 0976-
January

98
96
94
92
90 Series1
88
86
84
S1

S2

S3

S4

S5

S6

S7

S8

S9

S10

S11

S12

S13

S14

S15
Figure 9. Graphical representation of recognition rate of sample words

5. CONCLUSION

In this paper, we have present a novel method for classification and recognition of
presented or
simple Hindi language two or three letter words without modifiers using fuzzy rule based
approach. Characters are first classified into seven different classes and then recognized class
wise. Few misclassification cases arise due to the presence of: some of the similar shape
es
characters such as & and & , and characters which can be written in more than one
way such as & . We have extracted features for all the basic characters of the language
for recognition process. Algorithms developed perform well and give fine results as the most
prominent features, such as vertical bar, curves, loops and lines, are used at classification and
,
recognition stage. Experimental results verify the significance of the proposed system with
of
92.02% recognition rate. Fuzzy logic performs better than other methods as it can deal with
imprecise, incomplete and vague data efficiently without losing any important information. In
future, we will work to achieve better results and to improve the recognition rate by
emphasizing more on characters having similar shape such as and on Hindi words with
modifiers.

REFERENCES

Journal Papers:
[1]. N. Arica and F.T. Yarman-Vural, An overview of character recognition focused on
Yarman
off line hand writing, C99-06-C-203, 2000,IEEE.
C99
[2]. I.K. Sethi, and B. Chatterjee, Machine recognition of constrained hand printed
rinted
Devnagari,
pattern recognition, vol. 9, no. 2, 1977, pp.69 – 75.
,
[3]. R.M.K. Sinha and H. Mahabala, Machine recognition of Devnagari script, IEE IEEE
Trans. System, Man Cybern. 9,1979, 435-441.
435
[4]. S. Palit, B.B. Chaudhuri, P.P. Das, B.N. Chatterjee, Pattern Recognition, Image
Processing and Computer Vision, Narosa Publishing House, India,1995,163
1995,163-168.

356


[5] R. Plamondon and S. N. Srihari, “On-line and off-line handwriting recognition: A
comprehensive survey”, IEEE Trans. Pattern Anal. Mach. Intell., vol. 22(1), 2000,
pp63–84.
[6] M. Hanmandlu, O.V. R. Murthy and V. K. Madasu, fuzzy model based recognition of
handwritten Hindi characters, 0-7695-3067-2/07, 2007,IEEE.
[7] P. Mukerji and P.P. Rege, Shape Feature and Fuzzy Logic Based Offline Devnagari
Handwritten Optical Character Recognition, Journal of Pattern Recognition Research
4, 2009, 52-68.
[8] R.J.Ramteke, Invariant moments based feature extraction for handwritten Devnagari
vowel recognition, IJCA, ( 0975-8887) Vol 1 – No. 18., 2010.
[9] A. N. Holambe, R.C.Thool , Printed and handwritten character & number recognition
of Devanagari script using SVM and KNN, Int. Journal of Recent Trends in
Engineering and Technology, Vol. 3, No. 2, May 2010
[10] B. Singh, A. Mittal and D. Ghosh, An evaluation of different feature extractors and
Classifiers for offline handwritten Devnagari character recognition, Journal of Pattern
Recognition Research 2, 2011, 269-277.
[11] A. Pokhriyal and S. Lehri, MERIT: Minutiae Extraction Using Rotation Invariant
Thinning. International Journal of Engineering Science & Technology, vol. 2(7),
2010, 3225-3235.
[12] Primekumar K.P and Sumam Mary Idicula, “Performance Of On-Line Malayalam
Handwritten character Recognition Using HMM and SFAM” International journal of
Computer Engineering & Technology (IJCET), Volume 3, Issue 1, 2012, pp. 115 -
125, Published by IAEME

Proceeding Papers:
[12] P. Mukherji, P. P. Rege and L. K. Pradhan, Analytical Verification System for
Handwritten Devnagari Script. Proceedings of the Sixth IASTED VIIP, pp. 237-242,
Palma DeMallorca, Spain, August,2006.

Books:
[13] S.N. Sivanandam and S. N. Deepa, Principles of Soft Computing (Second Edition,
Wiley-India)
[14] R.C. Gonzales and R.E.Woods, Digital Image Processing (Second Edition, Prentice
Hall)

357

Fuzzy rule based classification and recognition of handwritten hindi

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Fuzzy rule based classification and recognition of handwritten hindi (20)

More from IAEME Publication (20)

Fuzzy rule based classification and recognition of handwritten hindi