ONLINE KANNADA HANDWRITTEN WORDS RECOGNITION USING DEEP LEARNING

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 3356
ONLINE KANNADA HANDWRITTEN WORDS RECOGNITION USING DEEP
LEARNING
Sangeetha G R1, Dr. Lakshman Naika R2
1P G Scholar, Department of Electronics and Communication Engineering, University BDT College of Engineering,
Davanagere, Karnataka, India
2Professor, Department of Electronics and Communication Engineering, University BDT College of Engineering,
Davanagere, Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The main objective of this paper is to recognize
online Kannada handwritten alphabet, kagunita and
ottakshara. Data collection is done using the graphics tablet
and stylus pen. Data collection consists of 25 samples of 49
characters of Kannada alphabet, 25 samples of kagunitaof all
the consonants, and 25 samples of 34 ottakshara. In total,
15,675 samples are collected. The main techniques used here
are Image Processing and Deep Learning. 80% of data is used
for training and 20% for testing. The classifier used in the
Deep Learning is CNN which yields 97.09% accuracy.
Key Words: Graphics Tablet, Stylus Pen, Deep Learning,
CNN, Image Processing
1. INTRODUCTION
Basically, Kannada is a major language in Karnataka.
Kannada is the official language in Karnataka and all the
notifications in government institutions are in Kannada
language. Recognizing the handwritten Kannada words is
the intention of this paper. Instead of using the keyboard
and typing Kannada by Nudi in laptop or computer, if
graphics tablet is used to write,whateverwegiveinput,that
will be automatically printed. This decreases the burden.
One who is unaware of keyboard, even he can writeand get
the written words in the standard text form.
Recognition is done by usingImage ProcessingandDeep
Learning. In Deep Learning Convolution Neural Network is
the best classifier that suits proper recognition of online
Kannada handwriting.
Data collectionof15,675samplesiscollectedtorecognize
alphabet, kagunitaaswellasottakshara.Thepenmovement
on the tablet helps to collect the dataset. Each data is stored
in different folders. Each folder consists of 25 samples of
Kannada character. Python language is used to write code
for data collection and data prediction.
2. LITERATURE SURVEY
Research is going on regarding recognition of Kannada
handwriting in online mode. Keerthi Prasad G, Vinay Hegde
[1] proposed online Kannada handwriting recognition. Data
is collected through thepenmovementonthegraphicstablet.
It can be writer dependent or independent. This work
focused on writer independent handwriting recognition.
Anirudh Ganesh[2] proposed for Kannada numbers using
CNN and DBN. The data collection is 74,000 and theaccuracy
of recognition is 97% in CNN and 98% in DBN. Salma Shofia
Roshyda, Tito Waluyo Parbuda[3] proposed for various
handwritingrecognition.AfterreviewingtheyconcludedCNN
is the best classifier which yields more accuracy. Gandhana
M[4] propsed for recognition of online Kannada characters
and numbers. B Vijay Kumar and Ramakrishna[5] proposed
for recognition of Kannada printed text. Rampalli and
Krishnamurthy[6] proposed for recognition of characters in
online and offline. Recognition in online is 11% more than
offline. Ragha L R et al., [7] the recognized forcharacterswith
recognition results 86% for vowels and 65% forconsonants.
ShashikalaParameshwarappaetal.,[8]proposedforKannada
handwritten character recognition with 99.09% accuracy
using SVM Classifier. Parikshith H et l.,[9] proposedamethod
in which any new style characters can be predicted.
Ganapatsingh G Rajputh et al., [10], proposed a method to
recognize characters with K-NN and SVM classifiers.
3. DATASET
The dataset is collectedthroughgraphicstabletandstylus
pen.to collect dataset weusecommandprompt.Thestepsare
as follows.
1. Create a folder and label it. Open command prompt by
the instruction ‘cmd’ the command window will be
displayed. Type, ‘code .’ the code will be opened, make
necessary changes regarding charactercountandfolder
name. Save it by pressing ‘S’.
2. Go to the command prompt window, type ‘activate
project’.
3. Next, type ‘python dataset_main.py’. It provides a check
window to collect dataset. Collect dataset of 25 samples
of a specified character. Press ‘S’, to save. Press ‘Q’ to
exit.
4. The data will be cropped and stored in the specified
folder.

5. Create new folder before collecting data for the
specified character and label it. Repeat step 3, until the
collection of data for all the characters.
6. Close the window after the collection of data.
(a) (b)
(c)
Fig-1 Dataset collection (a) alphabet, (b) kagunita
(c) ottakshara
Fig-2 Cropping of ottakshara
Fig-3 Dataset collected in folders
4. METHODOLOGY
This section describes the methods that are employed.
Data acquisition
The dataset is collected by using the Graphics tablet and
stylus pen. Once the graphics tablet is connected to laptop,
the connection is set up. Then thee dataset is collected and
saved in the specified folder. The size of the ottakshara
should be less than 50% size of the alphabet or kagunita. So
that it will be appended to the main character.
Pre-processing
Here the collected data is subjected to the resizing.
Whatever the size of input, the image will be resized to 52 X
52. Data collection is through online, so there is no need to
remove noise. The collected data is also subjected to
binarization. The characters appearing on the black colored
check window will be white. The data will besegmentedinto
three types, namely, character segmentation, word
segmentation, sentence segmentation.
Fig-4 Flow diagram of proposed methodology
The flow diagram of the proposed methodology is shown
above. It depicts the flow of the procsses that are involved.
It mainly uses the techniques of Image Processing and
CNN in Deep Learning which are best suited for this paper.
As online recognition is implementedthereisnooverheadof
noise removal, scanning the characters etc. Online
recognition is very interesting and can be widely used.
Online recognition gives better results and more accurate
when compared to that of offline recognition. There will be
no discontinuties because dilation,a morphological opration
is employed. The color images are converted to gray then to
black and white for preprocessing.
Segmentation
After data collection, the data need to be segmented. The
segmentation is of three types.
Character Segmentation
Word Segmentation
Sentence Segmentation
Let us take an example of sentences in which alphabet,
kagunita and ottakshara are present.

Fig-5: Segmentation of character, word and sentence
In the case of ottakshra, when the image is taken, the color
conversion of characters from color to gray and then from
gray to black and white conversion takes place.
Bonding boxes are applied. To identify the presence of
ottakshara, the height of the ottakshara should be less than
50% height of the original character. If not, there is no
ottakshara appended to the character.
In the above depicted figure 5, the character segmentationis
shown in thin red line, word segmentation is stored in the
thick red rectangular box. The sentence segmentation is
shown in blue color rectangular box.
In the output window whentheusergivestheinput,the
segmented characters, words and sentences are saved in a
folder. Let’s take an example. If the image is saved as 01-01-
01-0.png, it means that thee segmentation is done for first
sentence, first word and first character. 0 at th end tells that
there is no ottakshara. If 1 is present at the end it means that
there is a presence of ottakshara.
Each character extends pixel to 5 pixels towards right.
So that one word is recognized.
If 50 pixels are extended then it considers it as
sentence. Rectangular shape is used to recognize the
different words. To recognize the characters 5 pixels are
extended.
Image processing is the main technique that is
employed here to do segmentation. Cropping is done in this
process. The pixels are extended mainly to get the content of
data, so that no information is lost.
After character segmentation, if the segmentedcharacters
that are stored in the folders are not properly segmented, it
leads to wrong recognition of characters. Such data need to
be discarded.
If the segmented characters contain the proper character
that are written by the users, then such data is considered as
it gives the best accuracy in the output. Data segmentation
plays a very important role. If segmentationisdoneproperly
then the remaining next processes are easily carried out.
After data collection the characters need to be segmented.
The data collection contains 25 characters in each folder.
There will be 15,675 characters need to be segmented and
saved in the respective folders. The dataset includes these
characters which are shown below.
Fig-6 Kannada Alphabets
Fig-7 Kannada Kagunita
Fig-8 Kannada ottakshara

CNN
Using the CNN model, 80% of the data is used for training
and 20% of the data for testing. Among the available neural
network algorithms CNN is considered as the best. It
consists of three layers, namely 1. Input Layer 2. Hidden
Layers 3. Output Layer
Fig-9 CNN Architecture
Featureextraction andclassification aredoneusingCNN.
Hidden layers in CNN are: 1.Convolution Layer 2.Pooling
Layer 3.ReLU Correction Layer 4.Fully connected layer 5.
Softmax layer.
Labeling the data
Each collected data is assigned with a number. The
folder name that is assigned to the character is assigned to
the standard characters. So that computer recognizes
easily and efficiently.
Fig-10 Labeling the data
Training and Testing
In the training phase 80% of the collected data is used
and 20% of the data fortesting.Theaccuracyoftrainingand
validation increases and the training and validation loss
decreases.
Fig-11 Training and validation Accuracy
Fig-12 Training and validation loss
5. RESULT
Anaconda prompt isusedheretogetoutputwindow.There
are three windows which display the result. First window
enables user to give input and in same window output will
be displayed.
Fig-13 Input and output displayed in the same window
The second type of output window enables user to give
input in one ‘check’ window and after saving it by pressing
‘S’, the result will be displayed in the ‘result’ window.

Fig.14 Input and output displayed in separate windows
This window not only displays the combinationofkagunita,
it can also execute characters of alphabet. As we give some
space after eachcharacter, the space willalso be displayedin
the output ‘result’ window. This is illustrated inthefigure14.
Fig-16 Output displayed in notepad
The output window in which input and output are
displayed in the same window, the size o input window is
small. The input can be cleared by the user by pressing
‘clear’ button that is present below the output display.
The output cannot be edited. In other words, the output
cannot be copied to any other files. cut , copy, paste options
are not available.
The next type of output window in which input and
outputare displayed separately enables the user to writein
‘check’ window and displays the output in the ‘result’
window. Here, the input can be cleared by pressing ‘C’ on
the keyboard. If the input is properly writtentheoutputwill
be displayed as soon as the user press’ S’ on the keyboard.
This window also does not provide editing the output,
means increasing the font size or changing the font style or
copying the content etc.
The input window is somewhat wider, but the output
window is very small. It does not display if the user gives
input as a paragraph. But it will recognize the characters
perfectly. It does not need much time to receive input and
display it. The output style can be changed when the user
has some different font styles downloaded in the system.
In the third type, the input and output window both are
wide.
In the input window, where user gives the input, if he
writes anything wrong he may clear the screen. The input
can be cleared by using ‘C’ button on the keyboard. Then
user can give a new input in the input window. Thisprocess
goes on for every new input.
Once the input is fed to the input window, the user has to
press ‘S’. it takes some time to accept the input. Then the
user has to open the ‘data’ file in his system which is saved
before.
As soon as the file is opened, output will be displayed in
thee notepad. The user can change the font style, font size,
save the output. These changes can be made by the user in
this output window.
This window enables user to get the standard Kannada
text forcharacters, words, sentences, paragraph etc. So this
window is considered as the best among the three output
windows.
The main problem in this paper is huge data collection.
The user has to know how to hold the stylus pen, how to
hold the graphics tablet, where to start and where to stop
writing.

Fig-17 Graphics tablet connected to the laptop
It is easy for user one who knows all these techniques.
That person can collect data very fast. There will be no
overhead of discarding the characters once they are
cropped and saved in the respective folders. If any
characters are not cropped properly, or if two or more
characters are cropped in the same image, such images can
be discarded.
Fig-18 Cropped images are stored in the folders
While collecting dataset for ottakshara, the size of the
data should be small and many of the collected data is
discarded. Only the properly cropped characters are
considered. Again the remaining samples are collected.
While collectingdatasetforkagunita,somecharactersare
wider but they should be collected in the specified size. If
the supporting characters that are appended to the main
characters are missing it is also a overhead. Such data need
to be collected till the collection of 25 samples is collected.
The accuracy is not 100%. If more samples are collected,
then the accuracy will be increased. The input for some
characters needs to be written in a specified format. If the
input is not fed in that specified format, the detection of
such characters is wrong.
If any scratches are made to the words in the input
window will not be displayed in the output window. If any
symbols are fed by the user in the input window, those are
not executed in the output window, because the dataset
collection is composed of characters only.Thedatasetisnot
collection of symbols.
The output presented in the third type is veryhelpfuland
can be used to copy the output in the desired format in the
other files or folders. In Karnataka thisastheofficialnotices
are circulated in government institutions and offices is in
Kannada language, there will be variation in the font size in
the headingand content. Such modifications canbemadein
the output which is displayed in the standard Kannada text
format.
The output can be obtained at that instant as soon as the
user provides the input in theinputwindow.Thismakesthe
time to be saved. Instead of typing Kannada in laptop and
get it printed, if digital pad is connected and written it is
easy to use and easy to get the output.
Many people areunawareoftypingKannadainthelaptop
or in the computers. It is very useful for them. As the
technology is rapidly increasing from day to day everyone
has to know and implement and gt the work to b done in
easy way with less time. Input given to the system can even
be a song or proverb, it will be printed.
In this last method, thee background is white and the
input is black color. The output displayed is also in black
color. Color conversion process uses the Image Processing
technique. The thickness of input characters is more when
compared to that of the output. The thickness of output
characters can be increased by increasing the font size in
the notepad.
It enables user to give more input words and sentences,
and displays the output. It is the most beneficial and the
best window for user to give input and get the output and
increase the font size to the desired font size of the user.
The ottakshara will be appended to main character. If the
ottakshara is less than 50% of the main character.Training
accuracy is 99.23%, training loss is 0.0251%, testing
accuracy is 97.09% and testing loss is 0.210% by Tensor
Flow frame work.
As CNN classifier is used, it gives the more accuracy when
compared to that of other methods. CNN is used in the
Deep Learning that is employed here. It trains the system
as we train our mind.
CNN classifier is selected after reviewing so many
reference papers. Other classifiers are also used in the
reference papers but CNN gives the best result. So that the
prediction loss can be minimized and accuracy is more
during training as well as testing the data.

The ottakshara are collected in a manner that the main
character is not used along with it in the dataset. This
enables to get the output in such a manner that any
ottakshara can be appended to any consonant or any
kagunita . it gives a wide variety of characters, words and
sentences to be available in the standard Kannada text
form.
The output execution in the third type is more preferred
and wider output display enablestomakesomanychanges
like changing the font size and other benefits.
The output can be saved in the standard text and can be
opened at any time anywhere if it is saved in the laptop.
There is no need of internet to open the saved file in the
system. No need of other huge external devices to be
connected, it just needs a light weight digital pad to be
connected to the laptop.
The accuracy will be 100% if data is collected from many
people which consists a sample of 100 or 200. The chance
of getting the best accuracy is there for hugedata collection
of Online handwritten Kannada characters. There is no
overhead of misplacing and losing the data sitmayoccurin
the offline character recognition.
The available dataset can give a wide varietyofwordsina
standard Kannada text form. Online Kannada Words
recognition is very helpful and easy to implement in many
institutions and universities where Kannada is used the
most for circulars which are in Kannada language. As the
accuracy is more there is less overhead of wrong
recognition of the characters. The accuracy can be known
by ‘Tensor Flow’. The confusion matrix also tells about the
accuracy of the output for the proposed method.
The calculation can be made for outputdisplayintermsof
centimeter and pixels.
Fig-19 Output in the Microsoft Word
Let us consider the above figure.19.Theoutputofnotepad
is displayed in the Microsoft Word. The page size is A4. The
length and width of the A4 sheet is 21 X 29.7 cm. The
margin consumes 2.54cm in the left and 2.54cm in the right
side of the A4 size page.
There are 33 Kannada characters in the first line.
11Kannada words can be written in the above considered
example in A4 size page as illustrated in the figure 19. After
11 words, 12th word cannot be executed in the first line. It
has move to the next line of the sentence.
CALCULATION
Size of A4 sheet : 21 X 29.7 sq.cm
Margin left in the left side : 2.54cm
Margin left in the right side : 2.54cm
Total width left for margin : 5.08cm
Available space for executable line (A) :-
A = Total width – 2(margin)
A = Tw-2M
Width of A4 sheet : 21cm
Available width for the first line: 21cm-2(2.54)cm
Available width to for first line :15.92cm
2(margin) is considered in calculation. It is because, the
margin left in the left side as well as right side.
where, Tw is total width of A4 sheet
2M is margin in left and right
If the above calculation is written in terms of pixels,
Margin left in left side = 96 pixels
Margin left in the right side = 96 pixels
Total width for margin (2M) = 192 pixels
Width of A4 sheet (Tw) = 793.7007874 pixels
Width provided to print a line (W)
W= Tw- 2M
W= 601.700784 pixels
Width available for executing a line is given by the
formula:
Wc = ∑Wn + ∑Gw

Where, Wc refers to available width to execute a line
consisting of characters and space after each word.
Wn refers to width occupied by sum of characters
in the first line
Gw refers to gapleft after the executionofeachword
in a sentence in the first line.
6. CONCLUSION and FUTURE SCOPE
In this paper, online Kannada handwriting is recognized in
the online mode using the graphics tablet and stylus pen.
Data collection is very easy.Preprocessing,segmentation are
also important in this paper. Training and testing are done
using CNN classifier. Python is thelanguageusedto writethe
coding for data collection and data prediction. Image
Processing and Deep Learning techniques are used in the
proposed paper. This is very beneficial especially in
Karnataka as Kannada is an official language of Kannadigas.
In spite of wasting time in typing in the keyboard, the
output will be displayed in the standardformaswewritethe
words on the graphics tablet. The output is displayed in
three different windows, among them the last window isthe
best. Because the user can modify the font style and size,
copy the output, cut it or paste it anywhere, the user wishes
to do it in other folders.
There are many methods to get the characters to be
printed, among them this is the best method to be employed.
It is very helpful as the loss of data is less, it does not need
ink to write, it just uses thee stylus pen. As the technology is
improving this kind of techniques need to be employed to
get the work to be done even by a user who is unaware of
using ‘Nudi’.
This paper recognizes the Kannada characters in online
mode which is very beneficial an there is no loss of images
that are collected and no overhead of noise removing.
Nowadays, there is a huge scopeforhandwritingrecognition.
This paper explains the handwriting recognition of Kannada
language which is very helpful. It not only displays output in
words or sentence format, but also in paragraph format also.
It enables user to modify the font style and size according to
user wish.
There is no tension for the user regarding losing the data
collected, or misplaced or lost due to any reason. There is no
need of pen and paper. There is no need to scan the data for
saving and segmenting in the laptop or computers.
There is no need o battery for stylus pen. If the nibs are
flattened they can be removed and other nibs can be easily
inserted which takes only few seconds of time. Any number
of data can be collected without pen and paper. Thereisonly
one time investment for the graphics tablet. There is no
interference of noise which is the main problem in audio to
text recognition.
The execution of Online Kannada characters in the
standard text is easy and helpful formanypurposeswith less
time and less resources anywhere and anytime without
learning ‘Nudi’ which is popularly used by most of the
Kannadigas to get the standard Kannada text format in the
laptop or personal computers.
Online Kannada Recognition is very beneficial for many
Kannadigas present across the world. It recognizes the
characters, words, sentences and paragraphs and gives the
output at that instant only.
There is no need to depend on the people who can type
‘Nudi’ very well. One is unaware of Kannada typing can also
get the Kannada in the standard text format. Once the
process from data collection to testing iscompleted,theuser
can get any kind of output whether in the form of single
character or in the form of paragraph of many pages.
In future, characters or words with two or more
ottakshara can be executed which gives a completion of
executing any characters / words. More accuracy can be
made available in future. The biggest challenge for the
researchers in Kannada handwriting recognition in online
mode is to get 100% accuracy. No loss in accuracyneedto be
achieved. There are many words in Kannada with two
ottakshara, such words are included in the many sentences.
If those words are also included, then only there is
recognition of such words makes the words and sentences
meaningful and appropriate in Kannada language.
The execution of ‘O’ with any Kannada characters should
also be made to execute them in the standard form. Only the
consonants attached with ‘O’ are executing, but other
kagunita should also be made executed with ‘O’(anuswara).
Some characters which areusedinKannada but notincluded
in characters, kagunita and ottakshara need to be executed.
The above mentioned points can be implemented in offline
as well as online character recognition.Thereisa hugescope
in future for implementing the above mentioned points to
make a blended recognition in a single system which
recognizes the Online Kannada characters as well as the
offline Kannada characters so that whichever method is
convenient for user, user can use it according his choice.
Accuracy need to be increased by trying some other
different methods that are widely available. There is a
chance to get the 100% accuracy for the input provided by
the user in the future with no delay.
Instead of writing in the input window all the characters
can be made available in the portion of a window, in one
click on that character that character need to be printed in
the future.

Even with the less data collection theaccuracyoftheoutput
has to be made 100% in the future.So that the writer
independent system is implemented and can be used in the
future. So that there is no necessity of writing the words in
the input window as it is in the data collection. It recognizes
for any user handwritten characters in Kannada language
which is widely used in Karnataka by Kannadigas.
REFERENCES
1. Keerthi Prasad G, Vinay Hegde, “Recognition of Online
Handwritten Isolated Kannada Character using PCA
and DTW”, Interntional Journal of Recent Technolofy
and Engineering(IJRTE), 2019.
2. Anirudh Ganesh, Ashwin R Jadhav, K A Cibi
Pragadeesh, “DeepLearningApproachforRecognition
of Handwritten Kannada Numerals”, Springer
Internationl Publishing, 2018.
3. Salma Shofia Rosyda and Tito Waluyo Purboyo, “A
Review ofVariousHandwritingRecognitionMethods”,
Internationl Journal of Applied Engineering Research,
Vol. 13, 2018,pp 1155-1164.
4. Gandhana M H, “ Online Kannada Handwriting
Characters and Numerical Recognition using CNN
Classifier, IJERT, July 2021.
5. B VijayKumarndAGRamakrishnan,2004.RadialBais
Functionand Subspaceapproach for PrintedKannada
Print Recognition . Int. Conf. Acoustics, Speech Signal
Proc. 2004
6. Rampalli R Ramkrishnan, Angarai G, “ Fusio of
Complementary Online and Offline Strategies for
recognition of Handwritten Kannada Characters”,
Journal of Universal Computer Science, 17(1) . pp 81-
93. 2011.
7. Ragha L R, Sasikumar M, “Feature Analysis for
Handwritten Kannada Kagunita Recognition”,
International Journal of Computer Theory and
Engineering, IAC-SIT 3(1), pp . 1793-8201, 2011.
8. Shashikala Parameshwarappa, B V Dhandra, “Basic
Kannada Handwritteen CharacterRecognitionSystem
using Shape Based and Transform Domain Fetures”,
International Journal of Advanced Research in
Computer and Communication Engineering, Vol 4,
Issue 7, July 2015.
9. H Parikshith et al., “Handwritten Character
Recognition of Kannada Languag using Convolution
Convolution Neural NetworksandTransferLearning”,
IOP Conf. Series, 2021.
10. Ganapatsingh G Rajput, Rajeshwari Horakeri,
“Handwritten Kannada Vowel Character Recognition
usingCrackCodesandFourierDescriptors”,December
2011.
BIOGRAPHIES
Sangeetha G R is pursuing her
MTech in Digital Communication
and Networking, ECE Department
from UBDTCE, Davanagere, VTU,
India. Her major interest are in
Image Processing&DeepLearning.
Dr. Lakshman Naika R, completed
B.E (E&C) in Malnad College of
Engg, Mysore University, MTech
(Electronics Engg)inB.M.S.College
of Engg, VTU, PhD in Image
Processing, Jain University.

ONLINE KANNADA HANDWRITTEN WORDS RECOGNITION USING DEEP LEARNING

More Related Content

Similar to ONLINE KANNADA HANDWRITTEN WORDS RECOGNITION USING DEEP LEARNING (20)

More from IRJET Journal (20)

Recently uploaded (20)

ONLINE KANNADA HANDWRITTEN WORDS RECOGNITION USING DEEP LEARNING