Intelligent Recognition of Ancient Persian Cuneiform Characters

Fahimeh Mostofi

1,2

and Adnan Khashman

Intelligent Systems Research Center, Near East University, Near East Boulevard, Lefkosa, Mersin 10, Turkey

Department of Computer Engineering, Near East University, Near East Boulevard, Lefkosa, Mersin 10, Turkey

Keywords: Ancient Script, Persian Cuneiform Language, Character Recognition, Image Processing, Neural Networks.

Abstract: This paper presents an intelligent character recognition system based on utilising a back propagation neural

network model. The characters in question are unique and rare to be addressed in such applications. These

are the ancient Persian Cuneiform alphanumerical characters. The recognition system comprises firstly,

image processing phase where clear and noisy or degraded images of the ancient script are prepared for

processing by the neural model in the second phase. The importance of such application lies in its potential

to make translating ancient scripts and language easier, faster and more efficient. Experimental results

indicate that our proposed method can be further applied successfully to other ancient languages and may be

utilised in museums and similar environments.

1 INTRODUCTION

Being able to read and decipher ancient scripts has

always been of much interest and importance for

human-being. Many valuable historical secrets were

revealed by archaeologists through the time, made

us aware of the culture and civilization of our

predecessor.

The ancient near east was the home of early

civilizations such as Mesopotamia, ancient Egypt,

ancient Iran, Anatolia, Levant etc.

The Achaemenid Persian empire was the largest

that the ancient world had seen, extending from

Anatolia and Egypt across western Asia to northern

India and Central Asia.

Its formation began in 550

B.C (The Achaemenid Persian Empire, 2000).

The Old Persian language appears in royal

inscriptions, written in a specially adapted version of

cuneiform.

Old Persian cuneiform is a semi-alphabetic

cuneiform script that was the primary script for the

Old Persian language. Texts written in this

cuneiform were found in Persepolis, Susa, Hamadan,

Armenia, and along the Suez Canal (Kent

1950).They were mostly inscriptions from the time

period of Darius the Great and his son Xerxes the

Great kings of Achaemenid Empire (Kent, 1950).

Application of Artificial Neural Networks

(ANN) for pattern recognition and character

recognition has been more widely reported in

literature recent times. This has led to high

expectation of what neural networks can do for

different fields, especially fields where other

approaches have not been successful (Kashyap et al.,

2003).

Artificial Neural network is an information-

processing unit that is much inspired by the way the

human brain works. Brain can do some computation

(such as pattern classification and recognition) faster

than conventional computers (Patra et al., 2010).

Many researchers have been working on scripts

recognition for more than three decades.

Nevertheless, it remains to be one of the most

challenging problems in pattern recognition

(Kashyap et al., 2003).

Reading cuneiform symbols is an important

subject for understanding cuneiform tablets contents.

Where there are a great number of cuneiform tablets

representing valuable information of ancient history.

Yousif H. and et al proposed a method that used

the intensity profile curves for selected pixels in

images of Cuneiform text to differentiate between

them (Yousif et al., 2006).

In 2013 Naktal M. Edan proposed a method for

Cuneiform symbol recognition which was using a k-

mean algorithm for clustering similar Cuneiform

symbols and then classify symbols within the same

cluster by using a multilayer neural network (Edan,

2013).

In this paper we propose a multilayer feed-

119

Mostoﬁ F. and Khashman A..

Intelligent Recognition of Ancient Persian Cuneiform Characters.

DOI: 10.5220/0005035401190123

In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2014), pages 119-123

ISBN: 978-989-758-054-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

forward neural network based classification

technique for recognition of Old Persian cuneiform

characters.

The paper is organized as follows: Section 2

describes Old Persian Cuneiform alphabet and its

characteristics. Section 3 elucidates the general

architecture of old Persian Cuneiform classification

system. In section 3, image preparation step has

been described which explains how we supply

training and testing dataset. Section 4 explains

which image preprocessing methods have been used.

In section 5, we describe the design of multilayer

feed forward neural network for this application and

finally in section 6, results of the application are

discussed.

2 DESCRIBE OLD PERSIAN

CUNEIFORM ALPHABET

Most scholars consider old Persian writing system to

be an independent invention because it has no

obvious connections with other writing systems at

the time, such as Elamite, Akkadian, Hurrian, and

Hittite cuneiforms (Windfuhr, 1970). While Old

Persian's basic strokes are similar to those found in

cuneiform scripts, Old Persian texts were engraved

on hard materials, so the engravers had to make cuts

that imitated the forms easily made on clay tablets

(Kent, 1950).

The signs are composed of horizontal, vertical,

and angled wedges. There are four basic components

and new signs are created by adding wedges to these

basic components (Windfuhr, 1970).

These four basic components are two parallel

wedges without angle, three parallel wedges without

angle, one wedge without angle and an angled

wedge, and two angled wedges (Windfuhr, 1970).

Figure 1: Old Persian Cuneiform symbols consist of four

basic components (Kent, 1950).

The script is written from left to right (Daniels and

Bright, 1996).

Old Persian cuneiform alphabet set splits into set

of independent vowels (3 characters), constant

letters (33 characters), special signs such as signs for

country God and the king (8 characters), punctuation

(1 character) and numbers (5 characters).In

conclusion there are totally 50 separated signs for

old Persian cuneiform alphabet.

Figure 2: Old Persian Cuneiform Symbols (The Unicode

consortium v6.3).

3 THE PROPOSED SYSTEM

ARCHITECTURE

Figure 3: The proposed system architecture for Old

Persian character recognition.

NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications

120

Old Persian character recognition system consists of

five main steps: Image preparation, image pre-

processing, neural network design and training,

character recognition.

4 IMAGE PREPARATION

As we mentioned in previous section Old Persian

alphabet has totally 50 signs. In this paper we have

omitted special signs (such as king, country etc)

from the database. Instead in addition to the original

numbers which are 1, 2, 10, 20 and 100 we have

added numbers 3, 4, 30 and 40 in our dataset.

Thus finally our training set consists of 3

independent vowels, 33 constants, 9 numbers

(1,2,3,4,10,20,30,40,100) and 1punctuation character

which are totally 46 characters. Therefore the

training set in this paper consists of 46 jpg images

by size 64*64 pixels each is showing an old Persian

cuneiform alphanumeric character.

Figure 4: Old Persian characters which are included in the

training and testing dataset.

For creating a test dataset we have applied

Gaussian Filter in original images. Gaussian filtering

is used to blur images and remove noise and

details.in this paper we have used Gaussian filter to

Figure 5: Gaussian Function Distribution (Gaussian

Filtering 2014).

remove some details from images and make them

look blurred.

The Standard deviation (σ) of the Gaussian

function plays an important role in its behaviour.

The values located between +/- 3σ account for about

99% of the set. Larger values of σ produce a wider

peak (greater blurring) (Gaussian Filtering 2014).

We have tested the neural network by adding

different levels of noise to images. Setting σ to 3, 3.5

and 4 made the images blurred in three different

levels.

),(





eyxG





(1)

The equation (1) indicates Gaussian function

where x is the distance from the origin in the

horizontal axis, y is the distance from the origin in

the vertical axis, and σ is the standard deviation of

the Gaussian distribution (Nixon, Aguado 2002).

Figure 6: Demonstration of three different levels of

Gaussian noise applying to four Old Persian characters.

In figure 6 images in the first row indicates

firstly the original RGB image of character A and

then noisy images with Gaussian filter with





and then





3.5 and finally 



4.increasing value



produce more blurred images.the next third

rows are showing the same blurring process for

characters BA,SSA and CA.

5 IMAGE PREPROCESSING

By using image preprocessing techniques and thus

reducing image data, we can increase the speed of

learning by reducing computation expense.

In this section image preprocessing steps are

explained.

Firstly the RGB images are converted to

IntelligentRecognitionofAncientPersianCuneiformCharacters

121

grayscale, so that pixel values are in range 0 to 255

after conversion.

Secondly intensity adjustment technique has

been used for increasing image contrast, which maps

the image intensity values to a new range.In this

paper this task was accomplished by remapping the

data values to fill the entire intensity range [0, 255].

Thirdly for characters to be recognized efficiently

we need a suitable binarization algorithm which can

separate characters from the background accurately.

The binarization arithmetic is shown in Eq.2, where

),( jif is the original character image, ),( jif

the binarized image and T is the threshold. Otsu's

method has been used for determining T.













Tjif

jif

),(0

),(1

),(

(2)

Finally In order to make the training operation

faster we have complemented images.in this step the

images are binary so complementation is possible by

subtracting the image from 1.

Figure 7: Four steps of image preprocessing applied on

one original image in first row and the noisy image in

second row.

6 NEURAL NETWORK DESIGN

Among many neural networks arithmetic in

character recognition, back propagation (BP)

arithmetic is the most robust one.

The main principles of back propagation neural

networks can be concluded as following.

1) Input the character image into the BP neural

networks.

2) Calculate the Error Function by comparing the

recognizing result of the neural networks and the

known character result so as to adjust the value of

the connective parameters between layers(weights)

and make the output of the network more

approximate to the known result.

3) Train the networks with a series of known

images with the purpose of optimizing the

parameters of the networks (Y

ang,Yang, 2008).

In this paper old persian cuneiform characters were

classified by using back propagation neural network

The designed neural network consists of three

layers: Input Layer, hidden layer and output

layer.the number of neurons in input layer is 4096

which represents total number of pixels of a sample

image.the number of neurons in output layer is 46

which refers to the total number of Old Persian

character samples, and the number of neurons in

hidden layer is determined by experiment which is

44 in this neural network.

For both hidden and output layer, log-sigmoid

activation function has been used. As it has been

shown in table below Network Training parameter

epoch has been set to 5000 and training goal was

0.005.Training was stopped when number of epoch

is exceed or minimum error is meet. Momentum rate

was set to 0.45 while learning coefficient has been

set to 0.002.

The neural network is trained using 46 Old

Persian cuneiform character images. The 46 training

images are noise-free and are shown in examples in

Figure 3. The remaining images contain: 46 images

with





noise level, 46 images with

5.3





noise level and 46 images with





noise level .these are the testing images

which are not used in training set and shall be used

to test the robustness of the trained neural network in

identifying the characters.

Figure 8: Old Persian cuneiform character recognition

neural network topology

7 RESULT AND DISCUSSION

Table 1 lists the final parameters of the successfully

trained neural network, which learnt

after 243

iterations and within 13 seconds. The running time

for the neural network after training and using one

forward pass

was 0.07 seconds; this running time is

the time

required for the trained neural network to

identify

one character.

NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications

122

The old persian cuneiform character recognition

system was tested for three set of noisy images and

As it is shown in Table recognition rate for 46

images of training set was 100% correct

identification as it would be expected.

For the test set we applied three different levels

of Guassian filter

 By applying Guassian filter with





To 46 images 100% (46 out of 46) of Noisy

images were identified correctly.

 By applying Guassian filter with

5.3





To 46 images 93.4% (43 out of 46) of

Noisy images were identified correctly.

 By applying Guassian filter with





To 46 images 89.1% (41 out of 46) of

Noisy images were identified correctly.

Table 1: Neural network parameters.

Parameter Value

Input Neurons 4096

Hidden Neurons 44

Output Neurons 46

Learning Coefficient 0.002

Momentum Rate 0.45

Minimum Error 0.005

Maximum Iterations 5000

Iterations for training 243

Training Time(seconds) 13

Run Time(seconds) 0.07

Table 2: Neural network system's recognition rate.

Noise Level

3



5.3







TrainSet 100% 100% 100%

TestSet 100% 93.4% 89.1%

8 CONCLUSION

This paper presented Old Persian Cuneiform

character recognition system based on artificial

neural network.

The training database of this system consisted of

46 noise-free images of 46 alphanumeric old Persian

characters.

Applying three different levels of Guassian filter

with

3



, 5.3



and 4



to the original

images lead to creating a test dataset with totally 184

images.(46 noise-free images and 46 images for

each particular level of Gaussian filter.)

Training the neural network uses only 46 noise-

free images and testing the trained neural network

was performed using four different set of images

with different levels of noise.

The highest obtained rate for correct

identification of testing set Old Persian character

images was 100%. These images were not feed to

the neural network during training.

REFERENCES

Daniels, P., Bright, W., 1996. The World's Writing

Systems, Oxford University Press.

Edan , N., 2013. Cuneiform symbols recognition based on

k-means and neural network, fifth scientific conference

information technology, December 19-20.

Gaussian filtering lecture handouts of the University of

Auckland Department of Computer Science, 2014,

Accessed 21 April 2014, Available from:

<https://www.cs.auckland.ac.nz/courses/compsci373s

1c/PatricesLectures/Image%20Filtering.pdf>

Kashyap, k., Bansilal, Koushik, P., 2003. Hybrid neural

network architecture for age identification of ancient

kannada scripts,circuits and systems,2003.ISCAS

’03.Proceeding of the 2003 international symposium

on.(volum:5), 25-28 May.

Kent, R., 1950. Old Persian Grammar Texts Lexicon,

American Oriental Society, New Haven.

Nixon, M., Aguado, A., 2002. Feature Extraction and

Image Processing. Newnes, 1st edition.

Patra, P., Vipsita, S., Mohapatra, S., Dash, S., 2010. A

novel approach for pattern recognition, International

Journal of Computer Applications (0975 – 8887)

Volume 9– No.8, November 2010.

The Achaemenid Persian Empire,2000 , Accessed 21

April 2014, available from <http://www.metmuseum.

org/toah/hd/acha/hd_acha.htm >

The Unicode consortium, v6.3, Accessed 21 April 2014,

Available from <http://www.unicode.org/charts/PDF/

U103A0.pdf>

Windfuhr, G., 1970. Notes on the old Persian signs, Indo-

Iranian Journal XII/2, pp. 121-125.

Yang, F., Yang, F., 2008. Character recognition using

parallel BP network, Audio language and image

processing 2008.ICALIP., shanghai, July 7-9 .

Yousif, H., Rahma, A., Alani, H., 2006. Cuneifrom

symbols recognition using intensity curves, The

international Arab journal of information technology

vol.3,No.3 ,July 2006.

IntelligentRecognitionofAncientPersianCuneiformCharacters

123