See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339646668

Bangla Continuous Handwriting Character and Digit Recognition Using CNN

Chapter · March 2020

DOI: 10.1007/978-981-15-2043-3_60

CITATIONS READS

6 640

5 authors, including:

Fuad Hasan Shifat Nayme

Daffodil International University Daffodil International University
4 PUBLICATIONS 29 CITATIONS 4 PUBLICATIONS 29 CITATIONS

SEE PROFILE SEE PROFILE

Sheikh Abujar Md Mohibullah

University of Alabama at Birmingham Comilla University
102 PUBLICATIONS 936 CITATIONS 14 PUBLICATIONS 60 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Shifat Nayme on 19 June 2020.

The user has requested enhancement of the downloaded file.

 Bangla Continuous Handwriting Character and Digit
Recognition Using CNN

Fuad Hasan1, Shifat Nayme Shuvo1, Sheikh Abujar1,

Md. Mohibullah2, Syed Akhter Hossain3
1,3
Department of Computer Science and Engineering, Daffodil International University
2
Department of Computer Science and Engineering, Comilla University
1
{fuad15-9901, shifat15-9836, sheikh.cse}@diu.edu.bd
[email protected]
2
[email protected]

Abstract. There are several works in Bangla handwritten character recognition.

Here a new methodology proposed to recognize character from continuous
Bangla handwritten character. The system’s main components are
preprocessing, feature extraction and recognition. There is a strong possibility
which is found in Bangla words, characters are overlapped. This problem often
happens in handwritten texts like a consecutive character appear on another
character. When it comes to Bangla characters, segmentation becomes much
more difficult. To build an effective OCR system of Bangla handwritten text,
recognition of characters is important as much as segmentation of characters.
Here the main purpose is creating a system which takes continuous Bangla
handwritten text images as an input and first segments the input texts into its
constituent words and segments each word into individual characters. In this
present study, here we used Ekushnet dataset model which includes 50 basic
characters, 10-character modifiers, 52-frequently used conjunct characters and
10 digits. By using our algorithm, we are able to segments 95% words from text
and 90% characters from the words. Overall, in this present OCR system here
recognition and segmentation of characters from handwritten Bangla texts are
effectively dealing with the probable problems.

Keywords: preprocessing, segmentation, Removing Matra, modifier

detection, neural network, recognition

1 Introduction
In this globe every region has separate languages. As Bangla is one of the most
familiar dialects on the planet and approximately 220 million individuals apply
Bangla for their utterance and writing motive. In this present paper we work on
Bangla language. Therefore, recognition of Bangla continuous handwritten characters
has a great significance. This present article describes a system which can recognize
continuous handwritten Bangla texts and digits. OCR is a system which optically read
document to human readable form to machine understandable form. This system is
very popular in practical life application like language processing, library automation,
reading aid for blind, post office, banks, and government documents digitalized etc.
This system has a few modules, which are given below in a sequence such as
preprocessing, line segmentation, word segmentation, character segmentation,
character recognition. To obtain the best result Convolutional Neural Network (CNN)
has been used as a classifier for character recognition system.

2 Literature review
In 1870, a retina scanner which was invented by Carey [1], is an image transmission
system and it is the first character recognition system. There are two different types of
Bangla scripts are existing one is machine printed and another is Handwritten. In past
few years, there are many researches has been done in handwritten character
recognition in Bangla language. This research related works also achieves great
success. There are a couple of works are accessible for Bangla printed character
recognition system. Some significant work has been done in back years such as “A
complete Bangla OCR System for printed Characters” [2], “A complete OCR System
for Continuous Bengali Characters” [3], “An end to end System for Bangla Online
Handwriting Recognition” [4], “A hierarchical approach to recognition of handwritten
Bangla characters” [5], “ A complete printed Bangla OCR System” [6]. In all these
papers stated above shows various methodologies have been introduced by different
authors. Maximum works are done in printed continuous character recognition. But a
very few deals have been done with a complete OCR for handwritten continuous
character. From that standpoint, this paper is mainly proposed only continuous
handwritten character recognition. In this paper, Bangla continuous characters are
portioned utilizing some conventional approach just as some new philosophy.

3 Proposed methodology
In this paper to recognition of continuous handwritten character from sentences here
we present a new segmentation method for character. The main part done in the
character segmentation which has many phases that are following below.

Fig. 1. Illustrates OCR system

3.1 Prepossessing
In this section here we preprocess the input image. First step has been done by
converting the original image as grayscale image. Then removing the noise from the
image. Then convert the image as binary image to finding the foreground area from
the image. Elimination of unnecessary information as far as possible.
3.2 Line segmentation
Text line detection has been done by detecting the position between two consecutive
lines. Scanning the row horizontally if find white pixels that’s means it is a text.
Where the pixels are white considering as text figure. If a horizontal row is totally
black it is denoting a gap between two lines. That’s how a line can detected. In past
studies there are many works have been done in handwriting line segmentation in
different languages English [7], Hindi [8], achieve great success.

Fig. 2. Line segmentation

3.3 Word segmentation

The images are defined with binary pixel values, which is a binary image. Then
Compute the sum of pixels of each column of the image. If the sum of each column is
less than or equal to ten (as black pixel value is zero) than it is considered as a gap
between two words show in Fig 3. If fifteen connected gaps in both side of the word is
found than it is considered as a word. After finding the word remove unwanted
vertical and horizontal gaps (if sum of black pixels is zero) from the word shown in
Fig 4.

Fig. 3. Box Is Indicating Gaps Fig. 4. Word Segmentation

between two words

3.4 Character Segmentation

In this module find individual character from the segmented word. Here the main
problem occurs that, Bangla language has distinctive horizontal line known as
“matra” which connected all the character in a word. Bangla language also have
modifiers. Some occurs in the upper part and some occurs in the lower part of the
image. In every word there can be different types of modifier has been used. Some

Fig. 5. Before Matra elimination Fig. 6. After Matra elimination

word can have both the modifiers shown in Fig 7. To segmentation of character from
word, each word image is resized to h×w. Different types of word image can be found
(i) with no modifier, (ii) upper part or lower part modifier (iii) with both upper- and
lower-part modifiers. To identifying different types of word with different modifiers
there has been a flag indicator set. For different types of modifiers different method is
used to segmentation of character. Different types of word with modifiers described
below.

Fig. 7. word with both upper part and lower part modifiers
3.4.1 Word with no modifier
Considering that Bangla language has “matra”, the upper part of the word has
removed from the image. As the main body part of maximum Bangla character occurs
in the middle part. Considering that, the principle body of the word is taken form the
lower part of the image. To elimination of matra take the main body part by this
equation (Height-25). Here height is 100. From that image the connected white pixels
in y-axis is consider as an individual character shown in Fig 8. After indicating each
character removing the unwanted vertical gaps finally separates each character shown
in Fig. 9.

Fig. 8. Indicating characters Fig. 9. Segmented character with no modifier

3.4.2 Word with upper part or lower part modifier

For identifying the word with single modifiers, the lower part shown in Fig. 12(a).
and the upper part of image shown in Fig 12(b) is taken form the main image in Fig
10. Then split the image into three different parts shown in Fig 12(a) and Fig 12(b).
Each portion is here 1/3. If a word has an upper part modifier then at least the sum of
the black pixels of one portion (1/3) of upper part image will be zero. That’s how an
upper part modifier detection has been done. But in the lower part of the splitted
image, in three portion there is no zero (sum of black pixels) Shown in Fig 12(a). If a
word has a lower part modifier then at least the sum of black pixels of one portion of
the image of lower part will be zero. Shown in Fig 16. After identifying the word with
modifiers if a word has upper part modifier then the upper part is removed from the
image or if a word has lower part modifier then the lower part is removed from the
image. Then the same process will be use to find the connected point as section 3.4.1.
to segmentation of each character.

Fig. 10. Modifier above the matra Fig. 11. Upper part modifier detection

Fig 12(a). Lower Fig 12(b). Upper part Fig. 13. Segmented character with upper part

The lower part modifier is connected most of the time with the word. To remove the
lower part modifier from the word, calculate the starting white pixel point and ending
point vertically from the lower part of the image. After segmenting of each character,
the modifier is added after that character which belongs to this modifier.

Fig. 14. Lower part Fig. 15. Upper part no modifier Fig. 16. Lower part modifier
detected detected
3.4.3. Word with both upper- and lower-part modifiers
After discussing above two sections if a word has a both upper- and lower-part
modifiers then at least one portion (1/3) of both upper part and lower part has at least
one zero portion, Shown in Fig 17. For that reason, the word has both upper- and
lower-part modifiers. Then the same process will be used for character segmentation
as mentioned above in 3.4.1 and 3.4.2. section.

Fig. 17. Word with both lower and upper part modifier Fig. 18. Segmented character
and upper part modifiers

4. Classification and Recognition

For classification and recognition here, we use EkushNet [9] model. This model can
recognize the characters used in our daily life. It contains Bangla handwritten 50
fundamental characters, 10 digits, and 10 modifiers. It has also 50 compound
characters. This model use Ekush [10] dataset. With the help of CMATERdb [11]
dataset it is cross-validated, which recognition accuracy is 97.73%. Till now this is the
best accuracy for Bangla handwritten character recognition model.

Fig. 19. Architecture of EkushNet

5. Experimental Result
Few test samples of handwritten word image which is recognized by EkushNet
correctly in Fig. 20. By using this segmentation method from thousand of segmented
character about 70 percent character recognized properly. In Fig 21 some word image
which recognized falsely by EkushNet.

1 অা মা র 1 ো স গ ন ো র
2 অ িা প ন া র
2 বা া লা
িা ব ো শ ষ 3 ব া িা ধ ম া ন
3
4 বকাল 4 ব য় া ন্ধ র া

Fig. 20. Word image recognized correctly Fig. 21. Word image falsely recognize

Conclusion and future work

An approach for the character acknowledgment of Bangla manually written texts and
digits is presented in this paper from the beginning. The whole work describes
through point by point as main image converting in binary image, noise removing,
 line segmentation, word segmentation, removing matra, modifiers detection, character
segmentation and recognition. The hard work and the main challenge are in the
character segmentation portion, while two characters are simultaneously connected.
To get a great performance from an OCR system, it is dependent on proper feature
extraction of characters. For printed text the size and font are same but This is more
difficult task to segment characters from different handwritten text which is different
in shape and size. In future with more resources and a robust feature extraction of
characters can achieve a better result and recognition performance.

References
1. J. Mantas, An overview of character recognition methodologies, Pattern Recognition 19,
425-430 (1986).
2. I. S. Jacobs and C. P. Bean, “Fine particles, thin films and exchange anisotropy,” in
Magnetism, vol. III, G. T. Rado and H. Suhl, Eds. New York: Academic, 1963, pp. 271–
350.
3. Jalal Uddin Mahmud, Mohammed Feroz Raihan and Chowdhury Mofizur Rahman, “A
Complete OCR System for Continuous Bangla Characters", IEEE TENCON-2003:
Proceedings of the Conferenceon Convergent Technologies for the Asia Pacific, 2003.
4. S. Bhattacharya, D. S. Maitra, U. Bhattacharya, S. K. Parui, "An end-to-end system for
Bangla online handwriting recognition", 15th Int. Conf. on Frontiers in Handwriting
Recognition, pp. 373-378, 2016.
5. S. Basu, N. Das, R. Sarkar, M. Kundu, M. Nasipuri, and D. K. Basu, “A hierarchical
approach to recognition of handwritten Bangla characters,” Pattern Recognit., vol. 42, no. 7,
pp. 1467–1484, Jul. 2009.
6. B. B. Chaudhuri, U. Pal, “A complete printed Bangla OCR system,” Pattern Recognition,
vol. 31, pp. 531–549, 1998.
7. G. Louloudisa *, B.Gatosb,1, I.Pratikakisb,1, C.Halatsisa (2009). Text line and word
segmentation of handwritten documents.
8. G. S. Sindhushree, R. Amarnath and P. Nagabhushan (2019), Entropy-Based Approach for
Enabling Text Line Segmentation in Handwritten Documents
9. AKM Shahariar Azad Rabby, Sadeka Haque, Sheikh Abujar, Syed Akhter Hossain,
EkushNet: Using Convolutional Neural Network for Bangla Handwritten Recognition,
Procedia Computer Science, Volume 143, 2018, Pages 603-610, ISSN 1877-0509
10. Ekush: A multipurpose and multitype comprehensive database for Online Off-line
BanglaHandwritten Characters, Website: https://github.com/shahariarrabby/Ekush. Last
access:20 Jun. 18
11. R. Sarkar, N. Das, S. Basu, M. Kundu, M. Nasipuri, and D. K. Basu, “Cmaterdb1: a
database of unconstrained handwritten Bangla and Bangla– English mixed script document
image,” International Journal on Document Analysis and Recognition (IJDAR), vol. 15, no.
1, pp.71–83, 2012

View publication stats