A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different

Application Domain
Section A-Research paper

A Novel Preprocessing Technique for the Preservation of

Tamil Brahmi Letters on Ancient Inscriptions in
Different Application Domain
Poornimathi.K1,Muralibhaskaran.V2,Priya.L3
Assistant Professor(SG) 1,
Professor2,3,
RajalakshmiEngineering College1,2,3
1
[email protected] ,
2
[email protected] ,
3
[email protected]
+91 9790719507

Abstract: Inscriptions play a crucial role in preserving in this world for more than 5,000 years. Based on the market
historical, cultural, and linguistic information. The identification survey, the Tamil-only periodicals are numbered more than
and analysis of patterns in Tamil letters found in inscriptions 1800. These periodicals have been classified into three
provide valuable insights into the evolution of the Tamil categories based on the thorough analyses of grammatical and
language and its script. However, manual analysis of these
lexical variations, namely: Old Tamil (from about 450 BCE
inscriptions is time-consuming and prone to errors. In recent
years, deep learning techniques have shown promising results in to 700 CE), Middle Tamil (700–1600), and Modern Tamil
pattern recognition tasks, motivating the exploration of various (since 1600). Tamil-Brahmi, also known as Tamili, was a
strategies to identify the patterns of Tamil letters on variant of the Brahmi script in southern India. It was used to
inscriptions.This paper focuses on leveraging deep learning write inscriptions in the early form of Old Tamil. The Tamil-
algorithms for the automated identification of Tamil letter Brahmi script has been paleo graphically and stratigraphically
patterns in inscriptions. Firstly, a dataset of digitized Tamil dated between the third century BCE and the first century
inscriptions is collected, consisting of high-resolution images CE, and it constitutes the earliest known writing system
representing a wide range of letter variations. Preprocessing evidenced in many parts of Tamil Nadu, Kerala, Andhra
techniques are employed to enhance the quality and clarity of
Pradesh and Sri Lanka. Tamil Brahmi inscriptions have been
the images, removing noise and artifacts.Various deep learning
models, such as Convolutional Neural Networks (CNNs) and found on cave entrances, stone beds, potsherds, jar burials,
Recurrent Neural Networks (RNNs), are then trained using the coins, seals, and rings.
preprocessed image dataset. CNNs excel in extracting spatial
features from images, enabling the recognition of letter shapes From the 5th century CE onwards Tamil is written in
and contours. RNNs, on the other hand, capture temporal Vatteluttu in the Chera and Pandya country and Grantha or
dependencies within sequences of letters, aiding in deciphering Tamil script in the Chola and Pallava country. Vatteluttu is an
the structure and connectivity of the inscriptions. To improve ancient script that was used to write the Tamil language and
the performance of the models, data augmentation techniques
several other languages in South India. The term "Vatteluttu"
are employed to increase the dataset size and enhance its
diversity. However, preprocessing plays a major role in
means "rounded script" in Tamil, referring to the rounded or
sharpening the features present in the image. Hence, this paper circular shapes of its characters. This script has a rich history
addresses the preprocessing techniques such as Image Blur, and is of great importance in the development of writing
Binarization and Edge Detection with respect to inscription. systems in the region.It evolved from the earlier Tamil-
Preprocessing techniques were identified and tested with the Brahmi script, which was influenced by the Brahmi script of
inscription image. Based on the results and response time, it is North India. Vatteluttu script further developed into other
suggested that the Median filter with canny edge detection is scripts such as the Grantha script, which in turn influenced
working well for inscription images. After preprocessing the the modern Tamil script.Grantha script is an ancient writing
results have been tested with edge detection and it is found that, system primarily used for writing the Sanskrit language and
Median filter with canny edge detection gives best accuracy in
comparison with other algorithms. other languages of the South Indian region. It holds
significant historical and cultural importance, particularly in
the context of Tamil Nadu and Kerala. Both Vatteluttu and
Keywords— Character Recognition, Pre-processing, Image Grantha scripts are derived from the Brahmi script, but they
processing, Computer vision. belong to different categories and have distinct
characteristics. Vatteluttu is associated with Tamil and other
I. INTRODUCTION Dravidian languages, while Grantha script is associated with
Father of inscriptions is Samudragupta. However, Sanskrit and Tamil in a religious context.
Chandragupta Maurya was the one who has used inscriptions
to communicate with his subjects. In olden days inscriptions Tamil inscription images can be easily read and analyzed if
are the only format to communicate the human insights. The the preprocessing is done effectively. Hence, this paper
Dravidian literary languages include Tamil, Telugu, presented the various preprocessing techniques and its results
Kannada, and Malayalam. The eldest member of the for Tamil inscription images. The inscription images
Dravidian language family is the Tamil language. It has been

6373
Eur. Chem. Bull. 2023,12(10), 6372-6381
 A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

represent politics, religious, economy, location, tradition, addition, the researchers declare their intention to reconfigure
astronomy, history, culture and many others. the ANN's structure in future studies in order to improve the
system's predictive performance. [1]
Some of the ancient inscription images are shown below in
Fig 1 for a better understanding of inscription images.
Dhivya S et al (2021) suggested a few steps for efficient
character recognition. Training the CNN model from scratch
with a SoftMax classifier in a sequential model. Using
Mobile Net: Transfer learning paradigm from a pre-trained
model on a Tamizhi dataset. Building a model with CNN and
SVM. SVM for evaluating the best accuracy to recognize
handwritten Brahmi characters The accuracy of the trained
and tested Mobile Net model for the datasets of vowels (8
classes), consonants (18 classes), and consonant vowels (26
classes) is 98.1%, 97.7%, and 97.5%, respectively. Data
Collection Techniques: Approximately 1,600 engineering
students from the SRM Institute of Science and Technology
contributed 1,090,000 isolated samples to the character
database. The information was stored in CSV format. The
a.Ashokan Kalvettu b.Inscription at handwritten typefaces had a grid-like layout. Each grid cell
vikkiramankalam was segmented using the Hough Transform method. Using
( ) OPENCV, the image is segmented into distinct scripts. The
algorithm for character segmentation is used to extract
handwritten characters. Images of segmented characters are
classified into 209 classifications. To evaluate the dataset, a
CNN model and SoftMax classifier are combined to
recognize the script. This research may convince historians,
anthropologists, and scholars from other fields to utilize the
source material and gain a deeper comprehension of ancient
Tamil culture.[2]

The diverse thresholding techniques (Sukanthi et al., 2021)

The binarization of images of terrestrial and underwater stone
inscriptions is preceded by contrast enhancement and
followed by edge-based filtering that reduces noise and
sharpens edges. The modified bi-level thresholding (MBET)
c.Inscription at d.Inscription at Aanaimalai algorithm is proposed and compared to several existing
muthalaikulam ( ) thresholding algorithms, including the Otsu method, Niblack
( ) method, Sauvola method, Bernsen method, and Fuzzy C
means method. The proposed MBET algorithm, with its
adaptive local thresholding feature, is expected to minimize
Fig. 1 Sample Images of Tamil Inscription noise and extract the margins of the objects in both terrestrial
and underwater images flawlessly.[3]

II. LITERATURE REVIEW The technique (Brindha et al., 2021) for deriving the features
and converting the ancient Tamil script to its modern form is
In this section, a thorough discussion and analysis were proposed. The processed image is subjected to a new feature
conducted on the research work related to preprocessing extraction technique in which the system uses a chi-square
techniques utilized for inscriptions, as well as the technique test to determine whether or not all the zoning feature values
employed for the recognition of letters within inscriptions. are independent or dependent. Neural networks recognize the
characters using the extracted features. The featured image is
To identify prehistoric languages, signs, and fonts, Naresh et trained with NNTool, and the data are compared to the
al. (2022) used an approach that integrates an artificial neural database in order to recognize Brahmi characters. The
network (ANN) with the Opposition-based Grey Wolf recognized characters are converted into modern Tamil
optimization Algorithm (OGWA). The authors emphasize the letters, and the resulting HTML output is displayed. The
importance of the weights and connections between layers in image features are extracted using Zernike moments and
ANN system efficacy. This paper investigates the zoning features, two novel feature extractions. To ascertain
implementation of various optimization algorithms, including whether the values of the vectors are independent or
Opposition-based Grey Wolf optimization, Particle Swarm dependent, a chi-square test is conducted. The accuracy rate
optimization, and Grey Wolf optimization, to the ANN of 91.3% is achieved using the confusion matrix. By
system in order to determine these weights effectively. In employing sophisticated algorithms, the method can be

6374
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

expanded to larger datasets, online character recognition, and Tamil was developed. The effectiveness of OCR techniques
improved precision.[4] on ancient Tamil scripts can be enhanced through the
addition of more data. The development of a language parser
Neha Gautam et al. (2020) propose using a deep to aid in the segmentation of the digitized script and enhance
convolutional neural network with dropout to recognize the accuracy of the audio output is another possible extension
Brahmi words. In addition to proposing a DCNN for Brahmi of our work.[8]
word recognition, this study conducts a series of experiments
on a conventional Brahmi dataset. On a publicly accessible
Brahmi image database, the practical operation of this The proposed work (Merline et al., 2019) detailed the
method was systematically tested, obtaining a 92.47% historical events of the Chola period in the 12th century.
recognition rate by CNN with dropout. This study contrasted Using OCR technology, ancient Tamil characters carved into
the outcomes of various methods in order to make stones are identified. Ensemble learning and KNN are used to
recommendations based on parameter tuning. In addition, classify the characters, and Unicode is then used to match the
determining the optimal parameters for producing error-free classified characters. The images undergo pre-processing to
results is a research problem. Similarly, complex future tasks remove noise using median filters, segmentation using
such as character recognition of rotated, mirror-text, and bounding boxes, and extraction of features. The extracted
chaotic images could benefit from the extraction of novel features are applied to the Ensemble learning classifier. Using
features.[5] Unicode, the modern Tamil character is mapped. [9,13]

In this paper (Merline et al., 2020), training of an 18-layer A new system for improving stone inscription images is
CNN for 73 class character recognition problems was proposed by Durga et al., 2018.IBF is used to eliminate
conducted. This CNN architecture is trained to extract sample extraneous noise while maintaining character with edges. The
features using the ReLU activation function. CNN can proposed fuzzy system aids in predicting character and
autonomously learn a unique set of image-based features in a background pixel uncertainty. [11,12]
hierarchical fashion. We attained the Segmentation Rate and
Recognition Rate using our framework by mapping the From Summary of Literature is shown below for the user's
Ancient Tamil to Modern Tamil characters rate. The understanding.
proposed work concentrates on the Simple Convolution
Neural Network for image classification, which incurs less TABLE I
computational expense. The results demonstrate that CNN is Existing techniques- Summary
capable of obtaining good results on the Tamil Dataset
through supervised learning alone. We did not use an Author Year Technique Used Limitation
unsupervised, pre-trained network, despite the fact that we
believe it would have facilitated the work. [6] Naresh et. al ., 2022 Artificial Neural Better performance
Network (ANN) with by ANN
the Opposition-based
Using a Convolutional Neural Network, the proposed method Grey Wolf
(Suriya et al., 2020) is able to recognize characters in Optimization
challenging conditions where traditional character Algorithm (OGWA)
recognition systems fail, such as in the presence of low
resolution, substantial blur, low contrast, and other Dhivya S et. 2021 Convolutional Neural Identification of
al., Network (CNN) with Handwritten
distortions. The following significant challenges can be Mobile Net and SVM inscription only.
investigated further in the future: This work utilizes HP Labs
India's Isolated Handwritten Tamil Character data set. As a Sukanthi 2021 Edge-based filtering, PSNR and SD for
result, it has been determined that among the proposed et.al., modified bi-level the terrestrial stone
algorithms, various CNN models produce different results, Entropy thresholding surface-based
(MBET) images with 43%
and the one that provides the highest recognition accuracy is and 39% on an
superior. [7] average and for
underwater stone
The proposed work (Lalitha et al., 2019) concentrates on inscription images
49% and 39% on
enhancing optical character recognition techniques for 7th- to an average.
12th-century Tamil script. After the image is binarized using
the Otsu thresholding method, a two-dimensional Brindha et. 2021 Neural Network (NN) The 91.3%
convolutional neural network is defined and used to train, al., tool accuracy rate is
achieved using the
classify, and recognise ancient Tamil characters. The neural confusion matrix.
network is attached to Tesseract via the Python Pytesseract
library in order to implement optical character recognition Merline et. 2020 Convolutional Neural Improve the
techniques. This work incorporates Google's text-to-speech al., Network (CNN) efficiency by
voice engine to generate an audio output of the digitized text considering the
strokes, style and
as an added feature. This research endeavored to develop a size of characters.
universally applicable OCR system with audio output for the
ancient Tamil script. Using CNN and Image Recognition Suriya et. al., 2020 Convolutional Neural Handwritten
techniques, an operational system for modern and ancient Network (CNN) recognition only.

6375
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

Merline et. 2019 OCR and KNN The performance f. Normalization

al., metric of g. Image Scaling
segmentation &
recognition rates Out of the above Preprocessing techniques, the following
can be improved
techniques are being implemented and tested for Tamil
Naresh et. al., 2019 OCR, Advanced The geometric inscriptions.
Maximally stable features such as
extremal regions and edges and blob can a. Image Blurring
Affine Invariant be considered for
b. Binarization
Intensity Extreme efficient segment.
Based (AMSER) Light illumination c. Edge Detection
needs to be added.
a. Image Blurring:
Durga et. al., 2018 Modified Fuzzy 84.98% of
Entropy-based accuracy in
Adaptive extracting the
It is a technique used in image processing to reduce the
Thresholding characters from sharpness or clarity of an image intentionally. It involves
(MFEAT) with stone inscriptions applying a blurring filter to the image, which results in a
degree of Gaussian smoother or less detailed appearance. Blurring can serve
membership function
and iterative bilateral
various purposes, such as noise reduction, hiding sensitive
filter (IBF). information, or creating artistic effects.

There exists four different blurring techniques which are

widely used for inscriptions and are listed below:
III. METHODOLOGY FOR INSCRIPTION TRANSLATION
i) Average
A method has been presented in this paper from the ii) Gaussian
detailed study of literature present in this area. This method is iii) Median and
used for the translation of the ancient south Indian languages iv) Bilateral
into contemporary language using stone inscription images
from multiple geographical locations. Gaussian Blur : It applies a Gaussian distribution-based
kernel to the image. It provides a smooth blurring effect
while preserving the overall structure and reducing noise.

Median Blur : It replaces each pixel with the median value

of its neighboring pixels. It is effective for reducing salt-
and-pepper noise while preserving edges and fine details.

Bilateral Blur : It considers both spatial proximity and pixel

intensity similarity to perform blurring. It
preserves edges and details while reducing noise, resulting in
a smoother image with
preserved structure.

code snippet for preprocessing is given below for better

understanding.
Fig. 2 The proposed system's block diagram.
a. Image Blurring
The ancient Tamil letters from the stone inscription images
are being taken as the input. Noise levels present in the input import cv2
images are pre-processed using suitable algorithms and then from google.colab.patches import cv2_imshow
the image is getting binarized. Pre-processed image is from matplotlib import pyplot as plt
segmented and the features of the letters are extracted and img = cv2.imread('/content/kalvettu2.png')
based on the training model it has been classified and print('Original Image')
recognized as the Tamil letters which are in practice now. cv2_imshow(img)
# Gaussian Blurring
This paper mainly focuses on the preprocessing techniques gausBlur = cv2.GaussianBlur(img, (3,3),0)
involved in the inscription image translation are: print('Gaussian Blurring')
cv2_imshow(gausBlur)
a. Noise Removal cv2.waitKey(0)
b. Gray Scale Image # Median blurring
c. Thresholding medBlur = cv2.medianBlur(img,3)
d. Thinning and Skeletonization print('Median Blurring')
e. Skew Correction cv2_imshow(medBlur)

6376
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

cv2.waitKey(0) 1.Define the size of the local neighborhood (typically a

# Bilateral Filtering square or rectangular window) centered around each
bilFilter = cv2.bilateralFilter(img,3,25,50) pixel.
print('Bilateral Filtering') 2.For each pixel in the image:
cv2_imshow(bilFilter) a. Calculate the mean intensity of the pixel's local
neighborhood.
b. Determine the threshold value for the pixel by
subtracting a constant (often
referred to as the offset) from the
calculated mean intensity.
c. Compare the pixel's intensity with the threshold
value:
3.If the intensity is greater than or equal to the threshold,
assign it as foreground(white). and If the intensity is less
than the threshold, assign it as background (black).

Adaptive Gaussian Thresholding: Here, the threshold is

a. Input image b. Output image
determined by the weighted mean of the pixel's local
neighborhood using a Gaussian window.
Fig.3 Image Blurring
Steps to be followed:
b. Binarization 1.Define the size of the local neighborhood (typically a
The goal of image binarization is to segment an square or rectangular window) centered around each
image into foreground and background regions by assigning a pixel.
threshold value to each pixel. If the pixel's intensity value is 2.Calculate the Gaussian weight matrix based on the size
above the threshold, it is assigned the value 1 (white), of the local neighborhood. The Gaussian weight matrix
indicating it belongs to the foreground. If the intensity value assigns higher weights to the pixels closer to the center
is below the threshold, the pixel is assigned the value 0 of the neighborhood and lower weights to the pixels
(black), indicating it belongs to the background. It is used to farther away.
simplify an image and focus on specific regions or objects of 3.For each pixel in the image:
interest. Particularly useful in various image analysis tasks, a. Extract the local neighborhood centered around the
such as character recognition, document analysis, object pixel.
detection, and feature extraction. Some common methods for b. Multiply the pixel intensities of the neighborhood with
image binarization include global thresholding, adaptive the corresponding weights from the Gaussian weight
thresholding, Otsu's thresholding, and local thresholding matrix.
techniques. c. Calculate the weighted mean intensity of the
neighborhood.
Global Thresholding d. Determine the threshold value for the pixel by
It is a simple and widely used technique where a subtracting a constant (offset) from the calculated
single threshold value is applied to the entire image. Pixels weighted mean intensity.
with intensity values above the threshold are classified as e. Compare the pixel's intensity with the threshold value:
foreground, while those below the threshold are considered If the intensity is greater than or equal to the
background. This method assumes a bimodal histogram, threshold, assign it as foreground (white).
where there is a clear separation between foreground and If the intensity is less than the threshold, assign it as
background intensities. background (black).
Codesnippet:
If pixel_intensity >= threshold_value:
Set pixel_value = foreground_value #Adaptive Thresholding
else: import cv2 as cv
Set pixel_value = background_value import numpy as np
from matplotlib import pyplot as plt
Adaptive Thresholding img = cv.imread('/kalvettu2.png',
It is a technique that adjusts the threshold value cv.IMREAD_GRAYSCALE)
locally based on the properties of small neighbor hoods assert img is not None, "file could not be read, check with
within the image. It is useful when dealing with images with os.path.exists()"
uneven illumination or variations in object appearance. img = cv.medianBlur(img,5)
Common adaptive thresholding methods include: ret,th1 = cv.threshold(img,127,255,cv.THRESH_BINARY)
Adaptive Mean Thresholding: This method calculates the th2 =
threshold for each pixel based on the mean intensity of its cv.adaptiveThreshold(img,120,cv.ADAPTIVE_THRESH_M
local neighborhood. EAN_C,\
cv.THRESH_BINARY,11,2)
Steps to be followed:

6377
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

th3 = ret3,th3 =
cv.adaptiveThreshold(img,120,cv.ADAPTIVE_THRESH_G cv.threshold(blur,0,192,cv.THRESH_BINARY+cv.T
AUSSIAN_C,\ HRESH_OTSU)
cv.THRESH_BINARY,11,2) # plot all the images and their histograms
titles = ['Original Image', 'Global Thresholding (v = 127)', images = [img, 0, th1, img, 0, th2, blur, 0, th3]
'Adaptive Mean Thresholding', 'Adaptive Gaussian color = ['b','g','r']
Thresholding'] titles = ['Original Image','Histogram','Global
images = [img, th1, th2, th3] Thresholding',
for i in range(4): 'Image without Filter','Histogram',"Otsu's
plt.subplot(2,2,i+1),plt.imshow(images[i],'gray') Thresholding",
plt.title(titles[i]) 'Image with Filter','Histogram',"Otsu's
plt.xticks([]),plt.yticks([]) Thresholding"]
plt.show() for i in range(3):

plt.subplot(3,3,i*3+1),plt.imshow(images[i*3],'gray'
)
plt.title(titles[i*3]), plt.xticks([]), plt.yticks([])
plt.subplot(3,3,i*3+2),plt.hist(images[i*3].ravel(),2
56,color=color[
i])
plt.title(titles[i*3+1]), plt.xticks([]), plt.yticks([])
a.Input Image b.Output Image
plt.subplot(3,3,i*3+3),plt.imshow(images[i*3+2],'g
ray')
Fig.4 Adaptive Thresholding plt.title(titles[i*3+2]), plt.xticks([]), plt.yticks([])
plt.show()
Otsu’s Thresholding
This method automatically determines the
optimal threshold value by maximizing the between-
class variance. It computes the threshold that
minimizes the intra-class variance within the
foreground and background regions, resulting in a
better separation of objects from the background.
Otsu's method is particularly useful for images with
uneven illumination and non-bimodal histograms.

For each possible threshold_value:

Calculate the probabilities of pixels
being in foreground and background
classes
Calculate the mean intensities of the
foreground and background classes
Calculate the between-class variance
Select the threshold_value that maximizes
the between-class variance
Code snippets
import cv2 as cv
import numpy as np
from matplotlib import pyplot as plt
a. Input Image b. Output Image
img = cv.imread('/kalvettu2.png',
cv.IMREAD_GRAYSCALE)
Fig. 5 Otsu’s Thresholding
assert img is not None, "file could not be read, check
with os.path.exists()"
# global thresholding
ret1,th1 = Kernel or Filter
cv.threshold(img,127,255,cv.THRESH_BINARY) It is an essential parameter in image processing
# Otsu's thresholding algorithms, especially in operations like filtering and
ret2,th2 = convolution. The kernel, also known as a filter or a mask, is a
small matrix used to process the pixels of an image. It plays a
cv.threshold(img,0,255,cv.THRESH_BINARY+cv.T
HRESH_OTSU) crucial role in determining the nature and extent of the image
# Otsu's thresholding after Gaussian filtering processing operation applied. The importance of kernel size
blur = cv.GaussianBlur(img,(5,5),0) in image processing are Spatial Extent, Detail Preservation,

6378
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

Computational Complexity, Feature Extraction, Artifact within an image.Edges represent significant transitions in
Removal and Trade-off between Smoothing and Localization. image intensity and by detecting edges, subsequent
It is essential to understand the characteristics and algorithms can focus on analyzing and interpreting the
requirements of the image processing operation to select an structural information within the image.Edge detection is an
appropriate kernel size that balances the desired outcome important preprocessing step that sets the foundation for
with computational efficiency and the preservation of subsequent image analysis and interpretation tasks in image
relevant image details. processing pipelines. Most commonly used types of edge
detection methods:
import cv2
import numpy as np Steps Followed:
from matplotlib import pyplot as plt Sobel Operator:
image = cv2.imread('/content/kalvettu2.png') It is a popular edge detection method that uses a
# making filter of 3 by 3 filled with 1 divide by 9 for convolutional kernel to compute the gradient magnitude of an
normalization image. It approximates the gradient of the image intensity at
blur_filter1 = np.ones((3, 3), float)/(9.0) each pixel to identify edges. The Sobel operator is effective
# making filter of 5 by 5 filled with 1 divide by 25 for in detecting both horizontal and vertical edges.
normalization Canny Edge Detector:
blur_filter2 = np.ones((5, 5), float)/(25.0) It is a multi-stage algorithm that provides robust edge
# making filter of 7 by 7 filled with 1 divide by 49 for detection. It involves several steps, including noise reduction
normalization using Gaussian smoothing, gradient computation, non-
blur_filter3 = np.ones((7, 7), float)/(49.0) maximum suppression to thin edges, and hysteresis
thresholding to detect and link edges. The Canny edge
image_blur1 = cv2.filter2D(image, -1, blur_filter1) detector is known for its ability to accurately localize edges
image_blur2 = cv2.filter2D(image, -1, blur_filter2) and suppress noise.
image_blur3 = cv2.filter2D(image, -1, blur_filter3) Laplacian of Gaussian (LoG):
This method combines the Gaussian smoothing operation
plt.subplot(2,2,1),plt.imshow(image,cmap = 'gray') with the Laplacian operator to detect edges. It convolves the
plt.title('Original'), plt.xticks([]), plt.yticks([]) image with a Gaussian kernel to smooth it and then applies
the Laplacian operator to highlight regions of rapid intensity
plt.subplot(2,2,2),plt.imshow(image_blur1,cmap = 'gray') changes, corresponding to edges. The LoG method is
plt.title('Image_Blur(3*3)'), plt.xticks([]), plt.yticks([]) effective for detecting edges at various scales.

plt.subplot(2,2,3),plt.imshow(image_blur2,cmap = 'gray')
Table. II
plt.title('Image_Blur(5*5)'), plt.xticks([]), plt.yticks([])
Comparison of Image Blur with Edge Detection
plt.subplot(2,2,4),plt.imshow(image_blur3,cmap = 'gray')
plt.title('Image_Blur(7*7)'), plt.xticks([]), plt.yticks([]) Gaussian Blur

cv2.waitKey(0)
cv2.destroyAllWindows()

Median Blur

a. Input
Image b. Output Image

Fig.6 Kernel Size Bilateral Blur

c. Edge Detection

It is used to identify boundaries or edges between

different objects or regions within an image.It is performed
using various algorithms or operators, such as the Sobel
operator, Canny edge detector, or Laplacian of Gaussian
(LoG), among others. These algorithms analyze the gradients a. Input b. Output Image
or changes in pixel intensities to locate and highlight edges Image
Fig.6 Comparison of Image Blur with Edge Detection

6379
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

IV.RESULTS AND DISCUSSION Edge Detection Canny Edge Sobel Edge LoG Edge
The preprocessing of images was performed, and the (Bilateral Blur) Detection Detection Detection
response time for each step was measured. The obtained 0.94ms 2.17ms 1.28ms
results reveal important insights into the impact of
preprocessing techniques on response time in image
processing tasks. The analysis of the results highlights the
significant influence of preprocessing techniques on the
response time of image processing. Several key observations
can be made based on the data obtained. Firstly, it was
observed that the application of efficient noise reduction
algorithms during the preprocessing stage contributed to a
notable reduction in response time. By effectively reducing
noise and enhancing image quality, these techniques
streamlined subsequent processing steps, leading to faster
overall performance.
In addition, the incorporation of parallel processing
methodologies, such as multi-threading or GPU acceleration,
significantly improved the response time of the preprocessing
phase. By leveraging the computational power of multiple
cores or specialized hardware, the system was able to process
multiple images simultaneously, resulting in faster overall
execution.
Fig.7 Preprocessing Techniques with Response
However, it is worth noting that the achieved response time Time
improvements may vary depending on factors such as image
complexity, resolution, and hardware specifications. Further
evaluation and experimentation on a diverse set of images
and hardware configurations would provide a more
comprehensive understanding of the impact of preprocessing
techniques on response time. Over all, the results demonstrate
the effectiveness of the implemented preprocessing
techniques in reducing response time for image processing
tasks. These findings have important implications for
applications that require image processing in stone
inscription, enabling faster and more efficient analysis,
recognition, and manipulation of images.

Table III
Preprocessing Techniques with Response Time
2
Preprocessing Response Time in milliseconds
Techniques
Image Blurring Gaussian Median Bilateral
Blur=0.79m Blur=0.24m Blur=0.55m
s s s
Binarization Global Otsu's Adaptive
Thresholdin Thresholdin Threshold
g = 0.13ms g = 0.19ms Thresholdin Fig.8 Comparison of Image Blur with Edge
g = 16.17ms Detection Techniques
Kernel Size Filter Filter Filter
Size=3 Size=5 Size=7
V.CONCLUSION
0.33ms 0.43ms 0.75ms
This paper focuses on the utilization of various
Edge Detection Canny Edge Sobel Edge LoG Edge preprocessing techniques, which play a crucial role in
(Gaussian Blur) Detection Detection Detection enhancing the quality, readability, and interpretability of
0.97ms 4.53ms 1.74ms inscription data. Depending on the specific characteristics and
requirements of the inscription data, different preprocessing
Edge Detection Canny Edge Sobel Edge LoG Edge techniques offer distinct advantages. The selection of
(Median Blur) Detection Detection Detection appropriate preprocessing techniques for inscriptions depends
1.15ms 2.77ms 2.95ms on the specific goals, data characteristics, and subsequent
analysis tasks. In order to assess the response time of

6380
Eur. Chem. Bull. 2023,12(10), 6372-6381
A Novel Preprocessing Technique for the Preservation of Tamil Brahmi Letters on Ancient Inscriptions in Different Application
Domain
Section A-Research paper

preprocessing techniques, a table (Table III) is provided, [5] Neha Gautam,Soo See Chai , Jais Jose,Recognition of Brahmi words
byUsingDeepConvolutionalNeuralNetwork,https://www.researchgate.n
highlighting the efficiency of Median Blur in combination et/publication/341701111.
with canny edge detection. Median Blur is a powerful [6] Merline Magrina,Convolution Neural Network based Ancient Tamil
technique known for its effectiveness in image denoising and Character Recognition from Epigraphical Inscriptions-International
edge preservation. It offers noise reduction, edge Research Journal of Engineering and Technology (IRJET)-Volume: 07
preservation, and robustness against extreme values. Issue: 04 , Apr 2020.
Therefore, implementing preprocessing techniques in a [7] S.Suriya,Dhivya.S and Balaji.M,Intelligent Character Recognition
SystemUsingConvolutional Neural Network-EAI Endorsed
systematic manner significantly improves the legibility, Transactions-16October2020-DOI:10.4108/eai.16-10-2020.166659-|
quality, and usability of inscription data. This enhancement Volume 6 | Issue 19 | e5
enables various downstream applications, such as historical [8] Lalitha Giridhar Aishwarya Dharani Velmathi Guruviah,A Novel
research, cultural preservation, and automated content Approach to OCR using Image Recognition based Classification for
extraction. Furthermore, these techniques can be utilized for AncientTamilInscriptionsinTemples,https://arxiv.org/abs/1907.04917-
Jul 2019 .
accurate segmentation and recognition of letters in stone
[9] Merline Magrina M & Santhi M,Ancient Tamil Character Recognition
inscriptions, resulting in high accuracy rates. from Epigraphical Inscriptions using Image Processing
Techniques,Page40-48©MATJournals,Volume4|Issue2,2019.DOI:
http://doi.org/10.5281/zenodo.3253775.
REFERENCES [10] A.Naresh Kumar and Dr.G.Geetha,Character Recognition of Ancient
South Indian language with conversion of modern language and
translation, Caribbean Journal of Science-Volume 53, ISSUE 2 (MAY
[1] A. Naresh Kumar, and G. Geetha,Recognizing Ancient South Indian - AUG), 2019.
Language Using Opposition Based Grey Wolf Optimization ,Intelligent [11] K. Durga Devi & P. Uma Maheswari, Digital acquisition and character
Automation& Soft Computing,March 2022,Vol. 35,No. 3.DOI: extraction from stone inscription images using modified fuzzy entropy-
10.32604/iasc.2023.028349. based adaptive thresholding, © Springer-Verlag GmbH
[2] Dhivya S and Usha Devi G.TAMIZHİ: Historical Tamil-Brahmi Script Germany, part of Springer Nature 2018 images.Nov 2018.
Recognition Using CNN and MobileNet. ACM Trans. Asian Low- [12] Priya, L., Anand, S. Object recognition and 3D reconstruction of
Resour.Lang.Inf.Process.Vol.20,No. 3, Article 39 (June 2021), 26 occluded objects using binocular stereo. Cluster Comput 21, 29–38
pages.https://doi.org/10.1145/3402891. (2018). https://doi.org/10.1007/s10586-017-0891-7
[3] Sukanthi S, Sakthivel Murugan and S. Hanis, Binarization of Stone [13] Anand, S., & Priya, L. (2020). A Guide for Machine Vision in Quality
InscriptionImagesbyModifiedBilevelEntropyThresholding,Underwater Control (1st ed.). Chapman and Hall/CRC.
AcousticResearchLaboratory,WorldScientificPublishingCompany,Vol. https://doi.org/10.1201/9781003002826
20,No.6(2021)2150054(16 pages),DOI:
[14] https://www.tnarch.gov.in/epigraphy/inscriptions-tamil-script
10.1142/S0219477521500541.
[15] https://accounts.shutterstock.com-Photos of inscription
[4] S. Brindha and S. Bhuvaneswari, Repossession and recognition
system:Transliteration of antique Tamil Brahmi typescript-Current [16] “Tamil-Brahmi Kalvettu”-Sridhar-Published by Government of Tamil
Science,Vol. 120,No. 4,Feb 2021. Nadu Department of Archaeology

6381
Eur. Chem. Bull. 2023,12(10), 6372-6381