A Survey on Recognition of Ancient Tamil Brahmi

Characters from Epigraphy

A. Vidhyavani1, Dr.T. Manoranjitham2
1
Research Scholar, SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India. E-mail: vidhyava@srmist.edu.in
2
Assistant Professor (S.R), SRM Institute of Science and Technology, Kattankulathur, Tamil Nadu, India. E-mail: manorant@srmist.edu.in

The system which involves in character identification of Brahmi letters from epigraphy and converts to current Tamil character format.
Identification of earliest Tamil characters is one of the hardest part. If the letters are on the walls, it is still more difficulty in recognizing the
characters. Character recognition has reached near perfection in English and other language text. Identification of epigraphy brahmi characters
is very difficult because brahmi characters has been used from 3rd Century BCE to 4th Century CE. The accuracy level is not high in Brahmi
letters because of these reasons- First the consonants are similar to consonantal vowel characters, second the author doesn’t inscribe the Brahmi
letterings correctly and the fonts are inscribed in various styles and strokes. Only few people are known with ancient character if this remains,
all the valuable data given by our forefather will not be identified by the future generation. In this survey paper, we describe and compare many
techniques of identification through table.

Keywords: ICR, OCRs, NLP.

DOI: 10.47750/pnr.2022.13.S03.199

INTRODUCTION Categories of Tamil-Brahmi

Tamil is one of the eldest language in earth with rich poetry. Tamil-Brahmi has two types of development:
In olden days the poet in Tamilnadu used to write in palm
leaves, stones and inscriptions. Tholkappiyam which is Early (Tamil Brahmi)
written during 4th BC is the best example for palm leaf Early Tamil Brahmi has been used from 3rd Century BCE to
inscription which is the history of grammar book in Tamil. 1st Century CE. TB I and TB II vowel notational systems were
The ancient inscriptions contains valuable commentaries, in use.
some classics like saivam, vaishnavam, medical
commentaries, food, numerology, astrology, music, siddha, Late (Tamil Brahmi)
dance and many. Palm manuscripts mainly used for grammer,
It has been started from 2nd to 4th Century CE. The forms of
science, astrology, land registration which is donated by the
the characters gradually became more cursive which gets
king, historical, places.
converted into initial Vaṭṭeḻuttu characters. The schemes TBII
Old Tamil originated from Tamil brahmi script. Tamil script and TBIII has been used.
has been identified in wide range of area. Tamil Brahmi is the
Character recognition is the very complex work in pattern
source for old Tamil. The prediction of Tamil brahmi has
recocgnization system, because the researcher should know
been in use from 3rd century BCE. But the recent days
to separate the characters, and to identify the various fonts
evidences shows that it has been still back from 5 th century
with different style, categorize the character of same shape
BCE, but it do not have majority acceptance. Tamil-brahmi
and size.
had been in usage for many eras and gets adapted then gets
mutated into Vaṭṭeḻuttu from 5th Century CE.
OCR and ICR
The current Tamil is not directly derived from Tamil-Brahmi.
The current Tamil has been in vogue from 7th century OCR and ICR overall process looks same, but there is a main
Common era which is derived from Pallava grantha with differences between the two systems. OCR which is mainly
mixture of Vaṭṭeḻuttu. Upto 11th century CE Vaṭṭeḻuttu has to converts scanned images of text, even printed or
been used in Tamil Nadu. typewritten, and translates these photographs into machine-
encoded text. It is used as a file-keeping system for
businesses purpose and also used to post text through online.

Journal of Pharmaceutical Negative Results ¦ Volume 13 ¦ Special Issue 3 ¦ 2022 1273

 A. Vidhyavani, et al.: A Survey on Recognition of Ancient Tamil Brahmi Characters from Epigraphy

OCR is usually used to convert books or any large documents 94.10% and 90.62% for printed and handwritten Brahmi
into electronic files. character recognition respectively. As a whole, this method
Whereas ICR technically resembles as OCR but it is very offers a satisfactory success rate but the results could be
specific. An ICR is a method which is mainly to learn varies further improved by using NN and SVM techniques for
fonts and styles of handwritten characters. Using ICR, a classification of the Brahmi characters.
system can read handwriting and it is also recognize it to [3] Early Sri Lankan information’s are written on stone
increase accuracy and recognition. ICR is technically smarter surface are generally polluted with huge clatters such as
than OCR and also very detailed and more involved. cracks, scuffs, spaces, etc. The ancient letters can be
ICR is a subclass of OCR software, where the OCR software identified using precise method of alphabet fonts that tends to
is not set up to identify handwritten characters. Main automatic reading of ancient manuscripts by computers.
difference is that OCR is used for printed documents which This research work involved in automatic identification of
has been typed and translate into text and OCR text facilitates early Brahmi inscriptions by computer. Based on time period
copy paste. ICR focuses mainly on handwriting materials the shape of the letter gets changed. Also in a same period of
which has most difficult fonts than OCRs. time, one letter may get changed slightly. In this research, it
established automatic identification of ancient letters. They
proposed Modified correlation function method has been which
LITERATURE SURVEY is more sophisticated than the previous correlation peak
[1] The ancient history is being studied using the resource of method. Digital repository of Sri Lankan inscriptions is also
inscription and principles of development in the nation. trying to produce which has a role of automatic reading of
Recognizing and converting the early brahmi letters from the letters by computers.
temple epigraphy is one of the hardest part for current age [4] In this research twelve vowels from the palm leaf
group. Brahmi letters to Sinhala language automation is not manuscripts has been identified using B-spline curves. The
there. So manually they are translating the inscriptions. The uniqueness and robustness is the main advantage in B-spline
inscription is on the rock wall it is taking much time to covert curve. Each and every vowels in the Tamil character has more
the character into sinhala letters. This research major than one curves with various angle. The vowels can be
attentions on identification of early Brahmi letters inscribed identified by the combination of curves. Various narrators
between 3rd BC and 1st AD the time period. First, it eliminate inscription of same letter has been identified using this B-
the sound, sector the characters from the epigraphy pictures spline technique with the high level of accuracy than any
are taken and transform into binary format using Image other method.
Processing techniques. Then, it identify the broken letters and
convert that into correct Brahmi letters then recognize the B-spline curve method consist of three points P1, P2, P3
century of the inscriptions using CNN in machine learning. Where P1 is the first point, P2 is the middle point and P3 is a
At last, the Brahmi characters are coverted into Sinhala variable point. Then the variable point and the first point has
characters with the meaning of the letters using NLP. been connected and find the distance, then remove the point
P3, when it is less than a threshold value. Repeat until all the
The result of InceptionV3 model accuracy is 55.71. middle points are changed into P3.
DenseNet121 model accuracy is 60% and damage is 2.41.
Whereas the damage value of test set was greater than other The data set are collected from Aagama academy it is
model. Exceptions model accuracy is 46% and the damage executed in Python 3.0 and the different archives are tested
is 1.21. ReseNet50 model accuracy is 55.00% and the damage for the performance of vowel characters. It works based on
is 1.97. VGG19 model accuracy is 50.12% and the damage is the neural network with the support of B-spline curve [5] It
1.25. VGG16 model accuracy is 52.86% and the damage is involves in identification of Tamil letters from 9th to 12th
1.72. centuries First it pre-process the image then segmentation.
Based on threshold value the segmentation involves in
[2] It is a method for recognition of both handwritten and converting the images into binary image. Then the extraction
printed Brahmi characters which involve preprocessing, has been done by Scale Invariant Feature Transform
segmentation, feature extraction, and classification of Brahmi algorithms for each character to identify the correct letter.
script characters. The geometric method was used for feature
extraction into six different entities, followed by a newly Support Vector Machine classifier classifies the letters and
developed classification rules to recognize the Brahmi predicted by Trigram technique. Each recognized letter is
characters based on the features. The method obtains allocated with Unicode value and stored in the image
accuracy of 91.69% and 89.55% for handwritten vowels and warehouse
consonants character respectively and 93.30% and 94.90% It takes 50 pictures of each era. Othu’s Thresholding has been
for printed vowel and consonants character respectively. used for identification of character and Contorlet
Cropping, thresholding, and thinning method were used in the Transformation method for Recognition of characters.
preprocessing, Line detection and character detection method [6] In this research the text can be identified using RTI in
for segmentation before implementing feature extraction and which it has the characteristics of surface reflection. It choose
classify the characters. The accuracy of this method is one view over the object and it can identify the surface based

Journal of Pharmaceutical Negative Results ¦ Volume 13 ¦ Special Issue 3 ¦ 2022 1274

 A. Vidhyavani, et al.: A Survey on Recognition of Ancient Tamil Brahmi Characters from Epigraphy

on different lightings. Here the dome structure for RTI Shape Transformation
acquisition was done by 116 computers with controlled lights In shape transformation algorithm the transformation of
and a digital camera contains a stand. Every single set of character from image A to B has been done by
captured images were processed to generate the RTI image. 1) Some pixels are substitution.
In Cultural Heritage Imaging website RTI Builder open source 2) Some pixels are moved to the closest pixels in B, 3)
software is available. It can be worked remotely using web Some pixels are deleted due to deletion, insertion.
interface system. Two possible cases occur in this algorithm
Markout has been developed for RTI visualization tool, since i) Pixels remains in A;
for recognition and tracing phase of RTI visualization it is ii) Pixels which is not connected remain in B.
needed. iii) Pixels which is left in A were deleted. Moving the
pixel outside the character frame is eual to the cost of
[7] In order to maximize the speed and decrease the fault,
deleting a pixel A.
adaptive backpropagation learning model were proposed.
Network output contains error values and in each Iteration
hidden layers in the network is adapted. The change in values Defective images are identified using Hough Transform
and weights is noted in order to give efficient and less methods. In a parameter space the system of voting is taken
computation time. place, in which local maxima and accumulator space is
obtained. Straight lines can be identified in an image. Based
For identifing Tamil Letters ABP Model has been used 12
on that broken lines can be identified for feature selection.
vowel characters are selected. From that, 8 characters are
selected for training and 4 characters are selected for testing. [10] This study was created for recognizing cursive
C- Programming language has been used for coding and handwritten texts. It takes the handwritten image as input. For
verification. The character are taken as image format Then it feature extraction diagonal Feature Extraction is used, Euler
is digitized as binary format of 0 and 1 for the input. It is Number is used for classification. High accuracy rate is
segmented into 48 grids From 48 nodes, 8 were taken as achieved through Euler Number which is combined with
hidden nodes and 4 were taken as output nodes, which is like Diagonal Feature Extraction.
48-8-4. After conducting a set of tests on Fallaria system, this has
For training the network model all the nodes has been used been decided this system is more efficient than Fallaria
which is also called epoch Mean squared error is used for system. In this 100 characters has taken and got character
validation which can be identified by division of sum of accuracy of 88.78% and word accuracy of 50.4348%.
squared linear error in single node by double the maximum [11] This paper is based on the identification of special
number of node. The range of Network weights is between - characters and alpha numerals. Intelligence is needed for
3 to +3. Then the performance is tested with BP based on that identifying the characters. OCR works well in printed
Static parameter values and Elimination conditions was character but when it comes to handwritten or change in style,
determined. C surve measurement indicates the union of ABP font etc it becomes very difficult to identify.
model. For learning attributes and classification of labels Supervised
[8] This research is to recognize hand written Devanagari machine learning is used. By training the machine well it
letters by means of Deep Convolution Neural Network with shows accurate result.
transfer learning. It has been implemented in DenseNet, Vgg, Even for large amount of data ICR gives accurate result also.
AlexNet and Inception. It first classifies and send the data to excel sheet. The Using
Inception V3 and ConvNet. Inception V3 gives 99 % this method the result shows the accuracy of 95% for special
accurate result than any other methods in 16.3 minutes and characters and alpha numerals.
AlexNet performs with the accuracy of 98% at high speed of [12] This study involves in text identification of Brahmi and
2.2 minutes for sinlge epoch with a limited amount of data Vattezuthu characters from palm leaf and convert into Tamil
training the CNN model is very difficult task. digital text using the neural network and image zoning
So that, transfer learning has been used for solving these kind method. This algorithm contains image capturing, image
of problem. This transfer learning which transfer its preprocessing which consist of 4 process like.
knowledge to small data and gives accurate result. So CNN
take the same data for the beginning convolutional layers of i) Cropping
the network and it is going to train only the last few layers.
Usually the brahmi characters are inscribed with space and
[9] This research involves in digitizing the ancient Tamil some noise over the image. Cropping remove the unwanted
characters. For feature Extraction it uses Shape and Hough space and niose.
transform. Group Search Optimization, Firefly algorithm is
for feature selection. These algorithm is used for recognizing
the ancient Tamil script. ii) Segmentation
There are three types
a) line

Journal of Pharmaceutical Negative Results ¦ Volume 13 ¦ Special Issue 3 ¦ 2022 1275

 A. Vidhyavani, et al.: A Survey on Recognition of Ancient Tamil Brahmi Characters from Epigraphy

b) word v) Binarization
c) character In binarization 1’s are considered as a Dark pixel and 0’s are
considered as a Light pixel
iii) Resizing Than Data set training, Character Recognition, Unicode
Each character is in different size. So it resized all the Text, Retrieve from the database is done. Brahmi script
character in same size. accuracy rate is 91.57%. Vattezhuthu character accuracy is
89.75%.
So the Accuracy of Brahmi character is higher than
iv) Image thinning
Vattezhuthu.
A character can be converted in thin character.

AUTHOR METHOD/TECHNIQUES USED NUMBER OF ACCURACY

CHARACTERS TAKEN FOR
IMPLEMENTATION
K.A.S.A. Nilupuli CNN in deep learning, NLP Early Brahmi Characters, late Pretrained based model VGG 16-
Wijerathna brahmi characters 93.33% with loss-0.22, Pretrained
based model with different time
period, DenseNet121s
Neha gowtam Optical character recognition using Vowels and Consonants for Handwritten-90.62%, Printed
geometric method handwritten and printed character-9%s
characters
Nalin Warnajith Modified correlation function method, Early brahmi characters Early brahmi characters-55%
automatic correlation function method
Vowel 6 characters 85% accuracy
Suganya Athisayamani B-spline Curve Recognition

Manigandan T OCR and NLP techniques 50 Images of each centuries 72% accuracy
between 9th and 12th century.
Federico Ponchio Reflection Transformation Images 20000 characters 90% accuracy with only 10% loss
M. Sornam Adaptive Backpropagation Learning 45 epochs, 50 epochs and Identified 45 epochs in 56 milli
method 120 epochs second, 50 epochs in 67 milli
second, 120 epochs in 89 milli
second
Nagender Aneja Transfer learning for Deep 15 epochs Inception 99% accuracy in the first
Convolution Neural Network epoch at 16.3 minutes Vgg11 99%
accuracy in 45.6 minutes
AlexNet with 98% accuracyin6.6
minutes with 3 epochs
T.S. Suganya Shape and Hough transform, Group 9 characters each J48-93.33, KNN-91.75,NN-93.97
Search Optimization, Firefly contains 35 samples
algorithm
Yosuke R. Matsuoka Intelligent Character Recognition Word and charcters DFE with euler number-88.7
Diagonal Feature Extraction Vertical DFE without euler number-80.4
Feature Extraction VFC with euler number-87.2
VFC without euler number-79.1

Renuka Kajale intelligent character recognition, 200 samples taken 95% accuracy
supervised machine learning
E.K.VELLINGIRIRAJ Image Zoning method has been used Vowels, Consonants, Consonant Brahmi (vowels-93.45%,
vowels of both brahmi and consonants-92.75%, consonant
vattezhuthu characters vowels-90.24%)
Vattezhuthu (vowels-91.11%,
consonants-90.75%, consonant
vowels-89.12%) overall brahmi
character-91.57%, vattezhuthu
character-89.75%s

classification and final comparison. Optical character

CONCLUSION recognition plays a major role in identification of characters
The conclusion of this paper is, the recognition of characters and deep learning also helps to identify the character. The
from handwritten, palmleaf, and epigraphy. Most of the main aim of this paper is to identify method which gives
identification is based on image preprocessing, segmentation, accurate results from the epigraphy. Where the intelligent

Journal of Pharmaceutical Negative Results ¦ Volume 13 ¦ Special Issue 3 ¦ 2022 1276

 A. Vidhyavani, et al.: A Survey on Recognition of Ancient Tamil Brahmi Characters from Epigraphy

character recognition is mainly for handwritten and which is

an advance method of OCR. The ancient characters has been
identified in very few places in this literature review transfer
learning system very suitable for small dataset which train the
limited amount of data and gives accurate result.

REFERENCES
K.A.S.A. Nilupuli Wijerathna, Rashmi Sepalitha, Thuiyadura Indika,
Harshana Athauda, P.D. Suranjini, J.A.D.C. Silva, Anuradha Jayakodi,
Recognition and translation of Ancient Brahmi Letters using deep
learning and NLP, 2019 International Conference on Advancements in
Computing (ICAC), 978-1-7281-4170-1/19/$31.00 ©2019 IEEE.
Neha Gautam and Soo See Chai, Optical Character Recognition for Brahmi
Script Using Geometric Method, Journal of Telecommunication,
Electronic and Computer Engineering, e-ISSN: 2289-8131 Vol. 9 No.
3-11, https://www.researchgate.net/publication/322137737
Nalin Warnajith, Dammi Bandara, Sarkar Barbaq Quarmal, Masanori Itaba,
Atsushi Minato and Satoru Ozawa, Computer Analysis of
Photographic Data of Sri Lankan Early Brahmi Inscriptions, IOSR
Journal of Engineering (IOSRJEN), e-ISSN: 2250-3021, p-ISSN:
2278-8719 Vol. 3, Issue 1 (Jan. 2013), ||V3|| PP 44-49.
Suganya Athisayamani, Dr.A. Robert Singh,_, Dr.T. Athithanc,
Recognition of Ancient Tamil Palm Leaf Vowel Characters in
Historical Documents using B-spline Curve Recognition, Third
International Conference on Computing and Network Communications
(CoCoNet’19) Procedia Computer Science 171 (2020) 2302–2309,
Published By Elsevier.
Manigandan T, Dr. V.Vidhya, Dr.Dhanalakshmi V, Nirmala B, Tamil
Character Recognition from Ancient Epigraphical Inscription using
OCR and NLP, International Conference on Energy, Communication,
Data Analytics and Soft Computing (ICECDS-2017), 978-1-5386-
1887-5/17/$31.00 ©2017 IEEE.
Federico Ponchio, Marion Lame, Roberto Scopigno, Bruce Robertson,
Visualizing and Transcribing Complex Writings through RTI, 978-1-
5386-4385-3/18/$31.00 ©2018 IEEE.
M. Sornam, Muthu Subash Kavitha, M. Poornima Devi, An Efficient
Morlet Function Based Adaptive Method for Faster Back propagation
for Handwritten Character Recognition, 2016 IEEE International
Conference on Advances in Computer Applications (ICACA), 978-1-
5090-3770-4/16/$31.00©2016 IEEE.
Nagender Aneja_ and Sandhya Aneja, Transfer Learning using CNN for
Handwritten Devanagari Character Recognition, 2019 1st International
Conf erence on Advances in Information Technology, 978-1-7281-
3241-9/19/$31.00 © 2019 IEEE.
TS Suganya, Dr.S Murugavalli, Feature Selection for An Automated
Ancient Tamil Script Classification System Using Machine Learning
Techniques, PUBLISHED IN IEEE.
Yosuke R. Matsuoka, Gabriel Angelo R. Sandoval, Luis Paolo Q. Say, Jann
Skyler Y. Teng, Donata D. Acula, Enhanced Intelligent Character
Recognition (ICR) Approach using Diagonal Feature Extraction and
Euler Number as Classifier with Modified One-Pixel Width Character
Segmentation Algorithm, 2018 International Conference on Platform
Technology and Service (PlatCon), 978-1-5386-4710-3/18/$31.00
©2018 IEEE.
Renuka Kajale, Soubhik Das, Paritosh Medhekar, Supervised machine
learning in intelligent character recognition of handwritten and printed
nameplate, 978-1-5386-3852-1/17/$31.00 ©2017 IEEE.
E.K. Vellingiriraj, Dr.M. Balamurugan Dr.P. Balasubramanie, Information
Extraction and Text Mining of Ancient Vattezhuthu Characters in
Historical Documents Using Image Zoning, 2016 International
Conference on Asian Language Processing (IALP), 978-1-5090-0922-
0/16/$31.00_c 2016 IEEE.

Journal of Pharmaceutical Negative Results ¦ Volume 13 ¦ Special Issue 3 ¦ 2022 1277