Hindawi

Computational Intelligence and Neuroscience

Volume 2022, Article ID 3432330, 15 pages
https://doi.org/10.1155/2022/3432330

Research Article
A Deep Learning Approach for Recognizing the Cursive Tamil
Characters in Palm Leaf Manuscripts

Gayathri Devi S,1 Subramaniyaswamy Vairavasundaram,1 Yuvaraja Teekaraman ,2

Ramya Kuppusamy ,3 and Arun Radhakrishnan 4
1
School of Computing, SASTRA Deemed University, Thanjavur 613401, India
2
Department of Electronic and Electrical Engineering, The University of Sheffield, Sheffield S1 3JD, UK
3
Department of Electrical and Electronics Engineering, Sri Sairam College of Engineering, Bangalore City 562 106, India
4
Faculty of Electrical & Computer Engineering, Jimma Institute of Technology, Jimma University, Jimma, Ethiopia

Correspondence should be addressed to Yuvaraja Teekaraman; [email protected] and Arun Radhakrishnan;

[email protected]

Received 9 October 2021; Accepted 12 February 2022; Published 11 March 2022

Academic Editor: Syed Ahmad Chan Bukhari

Copyright © 2022 Gayathri Devi S et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Tamil is an old Indian language with a large corpus of literature on palm leaves, and other constituents. Palm leaf manuscripts were
a versatile medium for narrating medicines, literature, theatre, and other subjects. Because of the necessity for digitalization and
transcription, recognizing the cursive characters found in palm leaf manuscripts remains an open problem. In this research, a
unique Convolutional Neural Network (CNN) technique is utilized to train the characteristics of the palm leaf characters. By this
training, CNN can classify the palm leaf characters significantly on training phase. Initially, a preprocessing technique to remove
noise in the input image is done through morphological operations. Text Line Slicing segmentation scheme is used to segment the
palm leaf characters. In feature processing, there are some major steps used in this study, which include text line spacing, spacing
without obstacle, and spacing with an obstacle. Finally, the extracted cursive characters are given as input to the CNN technique
for final classification. The experiments are carried out with collected cursive Tamil palm leaf manuscripts to validate the
performance of the proposed CNN with existing deep learning techniques in terms of accuracy, precision, recall, etc. The results
proved that the proposed network achieved 94% of accuracy, where existing ResNet achieved 88% of accuracy.

1. Introduction across the nation in temples, libraries, and museums. An

average life of Palm leaves has 300–400 years, despite their
For future generations, knowledge from many academics conditioning capabilities and many processing, preserva-
was conserved in both written and oral forms. India went tion, and conservation procedures. Numerous palm leaf
through several ages of speech, drawing, and painting. manuscripts written three centuries ago are now in a state
Writing evolved from drawing through a sequence of of deterioration, and we risk losing the rich material
natural steps [1]. Different sorts of writing resources, such contained on them [5]. These valuable manuscripts provide
as leaves, wood, stones, barks, metals, and so on, were in-depth knowledge of Astronomy, Literature, Ayurveda,
found as writing increased. Palm leaves were frequently Construction, and Fine Arts written in many vernaculars
employed among all accessible materials due to their and scripts from diverse times. In ancient times, the only
abundant availability and ability to endure harsh preparing. method to repair and convey this precious material to
Palm leaf scripts are an important component of India’s future age group was to rewrite it. Digitization is being
written history [2–4]. Palm leaf manuscripts in various utilized to generate digital pictures of texts to preserve their
Indian languages and written scripts may be found all information as technology advances.
2 Computational Intelligence and Neuroscience

Tamil is among the world’s oldest and most recognized 2. Literature Review
languages, having a rich literary heritage. In the ancient
period, poets employed Palm leaves, notably in Tamil Tamil, one of the world’s oldest languages, is gaining
Nadu, to conceal information [6]. Sangam literature, popularity among scholars throughout the world due to its
masterpieces, Vaishnava, Saiva, medicinal works, gas- historical relevance and capacity to persist for centuries.
tronomy, astrology, Vaastu, gems, music, dance and Individual interests and ideologies of academics have shaped
theatre, and Siddha are all included in the ancient liter- study on palm leaves manuscripts in Tamil poetry over the
ature. The value of ancient medical texts in Tamil, as well as periods and in a variety of ways. The most prevalent topic of
the necessity to preserve them, has piqued the interest of research in the actual world is cursive character recognition
numerous academics in the recent decade [7]. The con- of ancient Tamil characteristics, and it focuses on the
served old medical writings in Tamil by saints such as documentation of Tamil characters to guarantee that a large
Agathiyar were generated for the first phase of a digita- amount of data is gathered. However, due to improper care,
lization process to preserve medical materials, and around palm leaf manuscripts have been destroyed, as numerous
10,000 manuscripts were successfully scanned [8]. For people held them by default. As a result, a rigorous ex-
digitizing historical documents, handwritten character amination of Tamil palm leaf texts, translation, and cata-
identification apps used three key methods: statistical, loguing of their publishing were necessary. Thousands of
structural or syntactic, and neural network-based palm leaf texts exist in the hands of institutes and specific
techniques. healers, but they must be digitized, and proper catalogues
Optical Character Recognition (OCR) is a method for were developed for the future. Similarly, numerous re-
converting various sorts of documents, such as scanned searches on the cursive handwriting recognition of the Tamil
papers. This method can be classified as either offline or language are included in the survey. There is, however, no
online. Offline handwriting recognition often uses a scanner universal technique for identifying cursive Tamil letters
to acquire the writing optically, and the entire writing is (with sufficient accuracy) in their whole. As a result, various
provided as an image [9–11]. Because the characters are approaches have been used in each stage of the recognition
identified as they are written, online handwriting recogni- process.
tion is also known as real-time recognition. Online OCR As such, this work is one of the first attempts to build
employs pen-based input devices to record the sequence of cursive training datasets by manual and automated segre-
coordinate points as the character is typed [12]. Handwritten gation of Tamil palm leaf scripts. Future researchers could
character recognition becomes complex due to various make use of the data set to develop expert systems that could
variations in the shape of characters, different writing styles, then be used for a range of purposes, such as a character
and overlapping and interconnection of neighboring letters. classification founded on the period they have changed and
It also depends on the individual because we do not write the the documentation of extinguished characters, and the
same character in the same manner. It is challenging to documentation of characters whose shape could have al-
create an OCR system with good recognition accuracy for tered. The study aimed to separate old Tamil palm leaf
Tamil script [13]. The primary goal of this method is to detect writings to provide a huge quantity of information about the
Tamil cursive letters in palm leaf manuscripts. This is ac- cursive Tamil character. There have been no previous in-
complished by categorizing the characters into relevant vestigations on the identification of cursive characters in
categories based on characteristics collected from each one. Tamil palm leaf texts. The ThimuNet of CNN is used in this
The steps below will assist in achieving greater accuracy. research project to classify cursive Tamil characters, and its
performance is confirmed. In the future, the produced
(i) Image Preprocessing training data might be utilized as input and fed into other
(ii) Feature Extraction deep learning systems for further categorization and finding
(iii) Segmentation of the cursive letters contained in manuscripts written in
(iv) Classification using CNN. Tamil.
Ali and Joseph [8] developed a CNN perfect for dis-
The leaf characters are segmented by using text line pensation real-time input pictures including Malayalam
slicing, where noises are removed, and morphological op- characters and the job of segmenting words and typescripts
erations are carried out in the work. Finally, the palm leaf from an image and attractiveness prediction using the CNN
characters are classified by using CNN model. The experi- model. The feature extraction job in this model is done
ments are conducted on the collected dataset by using implicitly in CNN by the gradient descent technique. This
various parameter metrics. The remaining paper is organized technique is efficiently utilized for digitizing Malayalam
as follows: Section 2 consists of Literature Review. The script, which comprises 36 consonants and 13 vowels, is
proposed methodology is presented in Section 3. The vali- approved out in stages, and has obtained an accuracy of
dation of projected methodology along with existing tech- 97.26 percent for the training dataset.
niques on the collected cursive Tamil palm leaves Balakrishnan and Pavithira [9] suggested a method for
manuscripts is presented in Section 4. Lastly, the research optimizing CNN using Simulated Annealing, demonstrating
work conclusion with its future work is given in Section 5. the efficiency of character appreciation. They discussed
Computational Intelligence and Neuroscience 3

several deep learning approaches, the capabilities of CNN, used to analyze the collected data. The study provides useful
and ways for training CNNs. According to the definition, information on the current collection of palm leaf organi-
“character recognition is the process of classifying and zations in Tamil Nadu, such as public libraries and private
distinguishing characters in an input picture and converting organizations, and nongovernmental administrations. It is
them to ASCII or any similar machine-accessible form.” The critical to have a sufficient quantity of knowledge on the
suggested approach assessed the OCR accuracy of multiple various approaches offered in Tamil Nadu’s palm leaf library.
language texts from numerous books, revealing that the This study indicated that manuscript conservation by dig-
accuracy of CNN by SA is higher than the unique CNN. italization is one of the most successful and helpful ap-
Hossian and Afrin [10] provided TEMPLATE MATCH- proaches; however, it is time-consuming and expensive.
ING, a method for extracting characteristics from pictures and Various knowledge schemes were documented in an-
identifying precise characters to generate written documents. cient India using palm leaf scrolls, according to Narenthiran
They claimed, “OCR operates similarly to the human capacity and Ravichandran [13]. The wisdom was primarily conveyed
to read.” The picture glance at by the scanner will be the source through the traditional Sanskrit language in various scripts.
for the OCR scheme in the same manner as an image viewed This study is based on five document libraries in Tamil Nadu,
by human eyes is the source for the brain in the natural which are a mix of libraries of varying statuses with superior
process.” They said that this methodology works by comparing manuscripts. This research concentrated mostly on the
the picture’s derived image characteristics and the template for cataloguing of manuscripts and the operations associated
each conceivable displacement. Techniques for word identi- with digitalized these archives. The study was based on a
fication and finding in images and videos may be classified as national goal, and the papers were recognized.
connected component-based methods and texture-based Sabeenian et al. [15] evaluated localized binarization
methods. We informed numerous ideas regarding Template techniques on Tamil palm leaf scripts, which were recorded
Matching, OCR, and the many stages required in Template to hold a huge quantity of info that is highly useful in ev-
Matching implementation. They tested the scheme by asso- eryday life. To protect the data on the palm, digital pictures
ciating the outcomes of taught datasets to those of unprepared of each leaf were evaluated and preserved. When necessary,
datasets, with accomplished datasets providing 100% accuracy. these pictures can be accessed. However, the storage volume
According to Baskar [11], the palm leaf was a frequent required for each leaf is considerable, which raises the cost.
writing material in primeval India, with the most basic These pictures must be binarized in order to acquire only the
surviving specimens. Furthermore, beginning in the sixth segmentation and storage. Several computer scientists have
century, the Palmyra palm was employed for document been stymied by poor reported picture binarization. This
production. Analysts and preservationists have recently been research focuses on palm leaf manuscript pictures that have
interested in the preservation of records and texts. There has been binarized using the Sauvola, Niblack, Bradley, and Nick
been some awareness in this respect, and their preservation binarization methods.
is important. There was a separate custom for palm leaf According to Kiruba et al. [16], one of the most essential
manuscripts, in which old texts that had become irreversible phases in manuscript recognition is character segmentation
had to be encrypted on fresh leaves, and the old ones were in Tamil palm script. Tamil is India’s most frequently used
disposed of in a river. Furthermore, religious laws prohibited script, and palm scripts contain consonants, modifiers, and
pundits from deporting the manuscripts to secure detention vowels. In addition, the accuracy of the recognition ap-
facilities. Lack of finances, insufficient apparatus, resources, proach is influenced by separate letters; therefore, suitable
and qualified staff in document sources are the mechanisms segments are necessary. This paper shows the image seg-
that impede effective care of palm leaf works. mentation from palm leaf manuscripts of Tamil handwritten
Sageer and Francis [14] suggested a digital library for letters. The technique consists of three steps: background
rare texts as a method for preserving palm leaves. The rate of removal using the Otsu algorithm to separate texts, character
deterioration to endangered records is higher in a variety of segmentation, and line segmentation. This paper proposes a
ways, and such leaflets are typically not constrained to the simple histogram-based method for segmenting Tamil palm
limited bounds of traditional library schemes. They are more script characters and discusses many obstacles to Tamil
commonly found in private groups, official records, and script segmentation.
institutional collections. The risk of harm is greater with As Ghosh et al. [17] note out that palm leaf works show a
private groups. Making a digital assortment for endangered significant role in India’s prized national legacy, particularly
materials significantly alleviates this situation. This research in relations of their extraordinary accumulation of famous
examines the different procedures of creating a digital ar- knowledge. Prior to the invention of paper, palm leaves were
chive for endangered materials and user attitudes toward most commonly utilized resources for encrypting messages
usability; it focuses especially on user attitudes toward the and literature. These books have survived the centuries and
digitization of palm leaf manuscripts. are still highly valuable today. The elements that deteriorate
According to Devika and Vijayakumar [12], traditional palm leaf documents have been investigated, with their
wisdom helps establish long-term interactions between conservation strategy, using both ancient and modern ap-
humans and nature. The purpose of this study was to dif- proaches. Because palm leaf works are important for
ferentiate the contents of palm leaves that can be digitized, knowledge conservation, the research suggested that they be
the details for digitization, and the various techniques of conserved as a first step. The research also suggested that the
scanning a palm leaf document. Statistical techniques were most essential task of today’s archivists and libraries is to
4 Computational Intelligence and Neuroscience

find efficient methods for the conservation of palm leaf Convolutional neural networks can be used to recognize
works. Tamil palm-leaf characters, as demonstrated by Sabeenian
Rendering to Challa and Mehta [18], there are numerous et al. [21]. Five layers of CNN were used in this study:
businesses in India dedicated to protecting the earliest palm convolution, pooling, activation, fully connected, and soft-
leaf scripts, owing to the importance of preserving the max classifiers. Scannable photographs of palm-leaf man-
valuable texts of knowledge. Both natural and artificial uscripts were used to compile the database of character sets.
processes influence palm leaves throughout time. With the There are 15 separate classes in the database, and each class
advancement of technology worldwide, this research has roughly 1000 samples. Generally speaking, CNN Clas-
attempted to digitize palm leaf scripts required in a uni- sifier’s recognition is determined to be between 96.1 and 100
versity library. One specific goal was to create an efficient percent. For each CNN layer, we extract a big quantum,
method of palm leaf texts for efficient information retrieval. resulting in an improved prediction rate.
Segmentation, picture augmentation, processing, acquisi-
tion, and compression are only a few of the techniques 3. Proposed Methodology
available. All of these technologies use distinct algorithms
that are effectively applied and produce the intended results. This section explains the workflow of the projected scheme,
Given that ideal identification rates are not feasible with where the four major steps are included. The first step is the
warped media and noise, selecting a proper image pro- collection of cursive Tamil palm leaves scripts, followed by
cessing approach is important. This research focuses on background removal, feature extraction, and classification as
image processing systems and algorithms utilized on palm the further steps. Figure 1 shows the workflow of the pro-
leaf texts projected by many scholars for effective data posed methodology. The following section describes the
recovery. significant things that are applied in the proposed model.
Vinoth et al. [19] investigated a method that aids in the
conservation of palm leaves. The Tamil language has been
identified as one of the world’s classical languages. Tamil 3.1. Dataset. The cursive Tamil palm leaf manuscripts are
Nadu was discovered to be a location where palm leaf collected from the online images and stored in the database
writings have been preserved. However, there are difficulties to document the target characters. In total, 100 images of
encountered with the protection of palm leaf writings. As a cursive palm leaf texts are collected. Figure 2 displays the
result, this research absorbed on the significance of con- sample collected cursive palm leaf texts. This data set is
serving palm leaf work by converting it to digital text format. enhanced and experimented by different methods, which are
Digital gadgets are currently used for this resolution of represented in the following section.
conservation. This digital technology aims to accurately
transform the characters and preserve the documents. 3.2. Preprocessing. Digitized palm leaf manuscripts image
Overall, the first step is to scan the palm leaf text and convert has dark leaf color as background and text in the foreground.
it to an image layout before saving it in a record. Following The presence of pickup noise in palm leaf manuscript images
that, the pictures should be converted into a digital text that happens while scanning or taking photo using the
format. digital camera can either be reduced by sharpening or using
The major goal of Sonam and Poornima’s [20] study any morphological methods. The sample image of the digital
was to create a system that allowed Tamil character noise occurred image is shown in Figure 3. The noise re-
identification from palm leaves, as well as caption moval is a necessary step to obtain useful information from
through the captured pictures and preserving them for digital text images. In preprocessing, the background has
future use. Confident training devices have been carried converted as black and foreground text as white in colors
out using a variety of approaches; however, the proce- from the range of 0 to 255 into 0 and 1 only so that the
dure of distinguishing the postures of Tamil characters is characters are very clear and easy for processing. The
difficult. It is also worth noting that the Tamil language background removal and morphological operations are the
is considered complex compared to other languages due methods to promote the images to be suitable for text line
to the existence of curves, slopes, pits, and twists, segmentation. According to the shapes of the leaf, the images
resulting in variations in different writing styles. Many are processed by the morphological operations, where
researches necessitate the adaptation of ancient Tamil structuring element is applied to an input image and ob-
characters to current characters to certify that the goal of tained the output image with same sizes. In MATLAB, a
developing computerized schemes for improving human structuring element called strel is included for this opera-
knowing comprehension is increased. This suggested tion, and finally, the output is obtained. By using this op-
research is thought to be useful for segmenting Tamil eration, noise and texture that are caused by simple
characters and keeping them in an orderly folder, as well thresholding are removed efficiently.
as for subsequent image processing. To enumerate the
statistical characteristics of the segmented characters,
GLCM matrix feature extraction is used. At this phase, 3.2.1. Background Removal. In most of these systems,
utilizing GLCM characteristics, the segmented Tamil binarization is the scanned gray level image (labeling each
characters are distinguished from those in palm leaf pixel) print or background in document image analysis.
manuscripts. Binarization is a process of assigning 0 s and 1 s using fixed
Computational Intelligence and Neuroscience 5

Input Cursive Tamil Palm

Leaf Manuscript

Pre-processing using
background removal and
Morphological operations

Feature Processing using

Text Line Slicing

Classification using CNN of

ThimuNet

Performance Validation

Figure 1: Working procedure of proposed methodology.

Figure 2: Sample cursive Tamil palm-leaf manuscript.

Figure 3: Sample noise occurred image before preprocessing.

6 Computational Intelligence and Neuroscience

threshold value. The fundamental idea of the fixed binar- 3.3.2. Space with Obstacle. The presence of an obstacle in the
ization method [21] is expressed in terms of relation. The space between the text lines can be categorized into two by
background of palm leaf manuscripts can be taken as black the length of an obstacle that helps decide whether it is
by 0 and the foreground text as white by 1. Here, T shows touching or overlapping text line. In Tamil character, an
global threshold value 50. After background removal, the obstacle is an important part of deciding the character. The
preprocessed sample output image is shown in Figure 4(b) length of an obstacle that extends and reaches the subse-
for the considered input image in Figure 4(a). quent text line is known as touching text lines (Figure 7). The
first line character “யு/you/” is touching with the next line
1, if f(x, y) ≥ T, character “த/tha/”. In Tamil, the text line segmentation is
g(x, y) � 􏼨 (1)
0, otherwise, complicated because if we ignore an obstacle of the character
“யு”, it becomes “ய”, and if we cut an obstacle in a fixed
where “x and y” is an input image sampled from the input length, the second line character “த/tha/” becomes “தி/thi/”.
data distribution, and “g” is the ground truth image cor- The overlapping text lines also have the same wrong pre-
responding to the input image. diction problem as touching text lines when we precede by
existing text line segmentation algorithms. The proposed
TLS solves the problem of touching and overlapping text
3.2.2. Morphological Operations. In this operations, dilation lines by fixing the cutting edge at the end of an obstacle and
and erosion are the fundamental operations. Addition of also prevents wrong predictions of the character.
pixels with the boundaries of text objects in an image is The purpose of text line segmentation is to precede the
known as dilation. Dilation is used to add pixels to the edges character segmentation. The touching and overlapping text
of regions or to fill in gaps in the image [22]. Erosion is the lines complicate the text line segmentation and make the
inverse of dilation. While dilatation raises the size of borders further process unproductive. An overlapping text line
and fills holes, erosion diminishes the size of boundaries and builds complication in-text line segmentation. An obstacle
enlarges holes. The reversal of this operation, i.e., extricating pervades the text zone of subsequent lines and mixed up
the pixels from the text object boundaries, is termed erosion with the character strokes that may precede wrong character
[20]. In order to process the text-image in palm leaf man- prediction or different from expected character.
uscripts, the pixels may be added or removed depending on
the size and shape of the text. In grayscale morphology, the
images are mapped into the Euclidean space or grid 3.3.3. Text Line Slicing (TLS). The proposed TLS line seg-
ES ∪ {r, −r}, where the grayscale erosion of the palm leaf mentation algorithm identifies an extension of character
image i by text boundaries b is given as in the following strokes using four variables such as Vertical space (Vs),
relation: Horizontal space (Hs), Vertical Track (VT), and Horizontal
∝ Track (HT). The variable Vs used to count zeros vertically to
(iΘb) � 􏽚 [i(x + y) − b(y)]. (2) know the stroke of a character exists in text line of binarized
yεB
Tamil palm-leaf manuscript image. The total columns count
of zeros is assigned to the variable VTand compared with the
threshold. The value 1 denotes that the space has no obstacle,
3.3. Feature Processing. The text line segmentation in and 0 represents an obstacle. The variable Hs is used to count
cursive Tamil palm leaf scripts is a Herculean task, and it the zeros horizontally and the total value assigned to HT,
influences till the end of the character recognition process. and then, it is compared with a threshold value. Three values
An absence of text line segmentation process is not possible are used to decide whether an obstacle is present or not; zero
for the successful character segmentation and character (0) defines the space between the character in text line; one
recognition in Tamil palm leaves. The TLS is applied on the (1) defines that the character has an obstacle; and two (2)
preprocessed binary palm leaf text images to segment the defines that the space is not found, which means that the
text lines. The new way of approach in-text line segmen- character exists concurrently.
tation of Tamil palm leaf images is to determine whether The obstacle creates touching text lines that can be
the obstacle is present between the text lines. Whenever the defined by Connected Component (CC). The connectivity of
strokes of the character exceed from the text zone and the character is calculated by the weight of the character
extend in the space between the lines, then it is considered using CC. The touching and overlapping text line characters
to be an obstacle in this case. The same is depicted in are considered single characters when they are connected to
Figure 5. each other (Figure 8). An algorithm implementation
(Figure 8(a)) proves obstacle identification in the space
between the text lines, defines the category of connected
3.3.1. Space without Obstacle. The TLS, the text lines of characters by connected component, and calculates the
Tamil palm leaf texts where the text line has enough space to weight for the character. When the minimum weight of the
the subsequent text lines or an elongation of character does character identifies the TLS, the algorithm implements
not reach the below text line as in (Figure 6, are considered as cutting edge (Figure 8(b)) to segment the connected text
space without obstacle or standard category. TLS can seg- lines. Cutting edge is a breaking point of touching characters
ment these text lines without any complication. in text lines. The CC provides continuation of the character
Computational Intelligence and Neuroscience 7

(a) (b) (c)

Figure 4: Preprocessed sample output image. (a) Input image. (b) Background removal. (c) Morphology.

Figure 5: Sample obstacle.

(a) (b)

Figure 6: Some of the sample text lines. (a) No obstacle, (b) obstacle not reached to the below text line.

touching ignoring cut in fixed-

characters obstacle length
Figure 7: Sample obstacle defined in the text line.
8 Computational Intelligence and Neuroscience

(a) (b)

Figure 8: Sample text line slicing. (a) Implementation on Tamil palm-leaf manuscript, (b) cutting edge.

strokes and also vertical stroke values. The minimum value where Lin is identified as an input length, Out is identified as
of the character stroke is known as the end of an obstacle that a length of the output, S is identified as a stride to filter slide,
must be fixed as a cutting edge for the text lines. The sample and Pad is identified as a padding.
images are shown in Figure 9. A CL comprises generally 3-dimensional input (Hin ×
Win × Cin ), where the height and width of the input are Hin
and Win and the channels of the input are Cin. The calcu-
3.4. Convolutional Neural Networks. Segmented and lation of the output function for each layer is exactly the
feature scaled images are given to the classifier model, which same. The CLs produce parameters, neurons, and various
is CNN. Despite their popularity, ANNs were incapable of connections, as mentioned earlier.
handling big datasets in recognition/classification respon-
sibilities. To overwhelm these problems, profound learning P � Wt + B, (4)
is a new learning machine paradigm. It is a stacked, mul- where B is a Bias, Wt is a weight of the CLs, and P is a
tilayered neural network. Earlier versions of the neural parameter. A CL weight can be calculated as
network, including the earliest perceptron, were superficial
with one input and one output layer and a hidden layer Wt � Cout × Hin × Win 􏼁 × Cin , (5)
between them. Each node layer in a deep-learning network
trains on a range of features depending on the earlier layer where Cout is output channels of the previous layer.
output. An ideal will be well-organized if hidden layers can CLs and PLs are used in basic CNN models to offer a
learn complex characteristics from seen data. Deep neural common architecture. CNNs use a succession of convolu-
networks outperform previously unknown data. CNNs tion operations to the input, coupled with/without pooling
utilize a variant of the multilayer perceptron to meet the and a nonlinearity activation function, before sending the
criterion of minimum preprocessing. CNNs are made up of output to the subsequent layer. The filters (F) are used in the
an automated feature extractor and a trainable classifier with CL to extract important characteristics from the input
several layers, such as picture for further processing. Each filter provides a unique
property for precise estimate. To keep the picture size the
(i) Convolutional Layer (CL), same, identical padding (zero padding) is used; then, valid
(ii) Pooling Layer (PL), padding is used since its assistance decreases the sum of
(iii) Fully Connected Layer (FCL). features. Each CL’s convolutional output may be written as
Win − F􏼁
Wout � + 1,
S
(6)
3.4.1. CNN Architecture. CLs and PLs are used in basic CNN
Hin − F􏼁
models, which offer a common architecture. CNNs apply a Hout � + 1,
sequence of convolution operations on the input, with or S
without pooling and a nonlinearity activation function, and where Wout and Hout are height and width of the output, and
then send the output to the next layer. The CL employs filters Win and Hin are height and weight of the input.
(F) to extract important characteristics from the input The weights and parameters change their values when
picture for further processing. Each filter provides a unique data flows over a deep network, occasionally making the data
property for accurate prediction. To maintain the image’s too huge or too small. This is called the problem of the
size, the (zero padding) same padding is utilized; otherwise, “interior covariate shift.” A modification in the distribution
its assistances decrease the sum of features. Each CL’s of the domain of a function is called a covariate shift. Each
convolutional output may be represented as layer entry is affected in deep networks by its parameters,
which means that a slight network change can affect the
Lin + 2 × Pad − F􏼁 whole network. Such internal layer modifications may cause
Out � + 1, (3)
S the deep network to experience an internal covariate shift
Computational Intelligence and Neuroscience 9

Figure 9: Illustration of the unique palm-leaf manuscript representing characters attained as an output with the typescripts.

issue. This problem is generally overcome by normalizing f(x) � max(x, 0), (11)
the data in each mini-batch. The general formulae of Batch
normalization (BN) are presented in the next section. It is where f(x) denotes the nonlinear activation function. The
used to allow every network layer to do learning more in- goal of FCL is to use the characteristics of the CLs and PLs to
dependently and used to normalize the output of the pre- classify pictures into distinct classes based on the training
vious layer. In addition, overfitting is minimized, because it dataset.
has a slight regularization effect. BN operates on 4D inputs, As previously mentioned, a computationally simple and
which may be represented as a small batch of 3D inputs. efficient cursive character specific neural network system is
During training, this layer retained a running evaluation of the aim. To accomplish this, a six-layer architecture is
its calculated mean and variance with a momentum of 0.1. proposed, with each layer essentially including CL > Batch
1 n Normalization (BN) > ReLU > Max Pooling Layer (Max-PL)
μ􏼂 xi 􏼃 � 􏽘x , (7) (see Figure 10). PLs are employed in the projected model to
n i�1 i
lower the spatial dimensions, which implies that they will
where μ[xi ] is the mini-batch mean value, xi is represented lessen the number of parameters inside the model, with a
as the value of the mini-batch elements, and process known as downsampling or subsampling. Following
each PL layer, BN is employed to deal with the interior
1 n covariate shift difficult. As the network becomes deeper, it
σ 2 􏼂 xi 􏼃 � 􏽘 x − 􏼂 x i 􏼃􏼁 2 , (8)
n i�1 i may become trapped in the saturation area, resulting in a
vanishing gradient issue/problem. To address this, the
where σ[xi ] is the mini-batch variance values: planned network employs Rectified Linear Unit (ReLU). The
xi − μ􏼂 xi 􏼃 complex patterns in the input data are learned by the net-
􏽢i � 􏽱
x ������ + ϵ, (9) work using activation function, and ReLU is considered as
σ 2 􏼂 xi 􏼃 an activation function in this model, where it does not
activate all the neurons at the same time. The goal of FCL is
􏽢 i is the mini-batch normalization:
where x to use the convolutional and PL features to categorize the
xi + β,
Yout � c􏽢 (10) input picture into multiple classes depending on the training
dataset.
where Yout is the mini-batch normalized value, and c and β The subsections that follow provide a brief explanation
are the learnable parameters: of the suggested CNN model construction.
10 Computational Intelligence and Neuroscience

FC Output

Conv1

Conv2
64 Conv3 Conv4
16 Conv5
64 32
8
32 16 8
4

128 256 512

64 Maxpool Maxpool Maxpool
32 Maxpool (2x2)
(2x2) (2x2)
Maxpool (2x2)
(2x2)
2048 Number of
Classes
Figure 10: Proposed modified CNN architecture—ThimuNet.

(1) ThimuNet Architecture. In the preceding sections, a (equation (4)) to reduce output errors. A concise archi-
fundamental introduction of CNN architecture was pre- tectural indication is as follows: 32C1 − BN − RELU − 2P2 −
sented. This section discusses the proposed CNN architec- 64C1 − BN − RELU − 2P2 − 128C1 − BN − RELU − 2P2 −
ture “ThimuNet,” which is meant to identify handwritten 256C1 − BN − RELU − 2P2 − 512C1 − BN − RELU − 2P2 −
cursive characters. The goal is to create a CNN model that 2048FC OUTPUT.where XCy denotes a CL with a total of X
can learn differentiating features for handwritten character kernels and a stride of y pixels in this representation, MP n
identification in a fraction of the time and space required by signifies a Max-PL 8 with a M × M pooling window and a
present models. Figure 10 depicts ThimuNet’s architecture, stride of n pixels, BN signifies a Batch normalization 18, and
which consists of six levels. The input image has a resolution NFC signifies a FC layer with n neurons.
of 64 ∗ 64 pixels. First, the supplied image is scaled to
(64 ∗ 64) pixels in size. The first layer then receives picture 4. Results and Discussion
pixels as input. Each CL alternates with a subsampling or
pooling layer, which takes the pooled maps as input. As previously stated, the objective of this study is to provide
CL, Max-PL, and FCL are denoted as Cx , Mx , and FCx a superior CNN model than the current models for the
correspondingly, in the following discussion, where x is the gathered cursive dataset. To build the suggested neural
layer index. The input for the first CL, C1, will be network design, PyTorch7 was used as the Python-based
64 × 64 × 1. The output of C1 ∗ (“C” after numerals rep- framework. On these gathered datasets, evaluation of many
resents layer number), ReLU, and batch normalization is modern models such as LeNet5, ResNet (18/34/50), AlexNet,
applied. As a consequence, 22 subsampling processes were DenseNet121, InceptionNet v3, and others was also con-
performed. M2 is a max-PL with 32 feature maps of 32 × 32 ducted to have a comparative analysis. All of these trials are
dimensions. Table 1 shows the calculation of trainable carried out on a system equipped with Intel Core i3 CPUs,
characteristics and trainable links. The C3 layer has 64 16 GB of RAM, and an NVIDIA graphics card with 4 GB of
feature maps. C3’s output feature maps are linked to M4. internal memory and 768 CUDA cores.
ReLU and BN are implemented after each CL layer. C5, C7,
and C9, like the previous convolutional layers, convolute the
prior feature maps. Max-PLs are linked to each subsequent 4.1. Result Evaluation of Proposed ThimuNet. This section
CL. validates the proposed CNN performance with other
The output volume will be 512 (2 × 2 × 512) after the final techniques in terms of various parameter metrics. The major
layer. As a result, they can extract characteristics that are parameters such as Precision, Balanced Classification Rate
more resistant to local changes of the input pictures. This FC (BCR), Recall, Sensitivity, Misclassification Penalty Metric
layer has been flattened to 1 × 1 × 2048, used to compute the (MPM), Specificity, Balanced Error Rate (BER), F-measure
class scores. This entire layer scheme is similar to a con- based on sensitivity and specificity, Peak Signal to Noise
ventional feedforward network by overlapping the map ratio (PSNR), and Distance Reciprocal Distortion (DRD)
element and kernel element input on each specific layer and can be computed for the Tamil Palm leaf manuscripts.
adding results together for the output to that particular layer However, in this work, only a few parameters are considered
after calculation. Gradients of error have been calculated via for the validation process; the reason is that it is a collected
backpropagation. Gradient descent was employed to update cursive Tamil Palm Leaf manuscript and cannot apply all
all network weights (equation (5)) and parameter values major parameters.
Computational Intelligence and Neuroscience 11

Table 1: Summarization of different parameters of the proposed ThimuNet.

Convolutional layer Parameters Neurons Connections Activations
Layer 0- INPUT
0 0 0 64 × 64 × 3 A � 12,288
(64 × 64) 3 channels
C � 32,76,800 Conv � 64 × 64 × 32
Layer 1- FILTER (5 × 5) Wt � 32 × (5 × 5) × 3 B � 32 (64 × 64) × 32
(64 × 64) × pool � 32 × 32 × 32
32 outputs P � (2400 + 32) � 2,432 N � 1,31,072
32 × (5 × 5) BN � 64 × 64 × 32 A � 294,912
(64 × 64) × Conv � 32 × 32 × 64
Layer 2- FILTER (5 × 5) Wt � 64 × (5 × 5) × 32 B � 64 (64 × 64) × 64
64 × (5 × 5) BN � 32 × 32 × 64
64 outputs P � (51200 + 64) � 51,264 N � 2,62,144
C � 65,53,600 pool � 16 × 16 × 64 A � 147,456
(64 × 64) × Conv � 16 × 16 × 128
Layer 3- FILTER (5 × 5) Wt � 128 × (5 × 5) × 64 B � 128 (64 × 64) × 128
128 × (5 × 5) BN � 16 × 16 × 128
128 outputs P � (204800 + 128) � 2,04,928 N � 5,24,288
C � 1,31,07,200 pool � 8 × 8 × 128 A � 73,728
(64 × 64) × Conv � 8 × 8 × 256
Layer 4- FILTER (5 × 5) Wt � 256 × (5 × 5) × 128 B � 256 (64 × 64) × 256
256 × (5 × 5) BN � 8 × 8 × 256
256 outputs P � (819200 + 256) � 8,19,456 N � 10,48,576
C � 2,62,14,400 pool � 4 × 4 × 256 A � 36,864
(64 × 64) × Conv � 4 × 4 × 512
Layer 5- FILTER (5 × 5) Wt � 512 × (5 × 5) × 256 B � 512 (64 × 64) × 512
512 × (5 × 5) BN � 4 × 4 × 512 pool � 2 × 2 × 512
512 outputs P � (3276800 + 512) � 32,77,312 N � 20,97,152
C � 5,24,28,800 A � 18,432
Fully connected layer
Wt � 512 × (2 × 2) × 46 B � 46
(FC layer)—46 class 0 0 A � 46
P � 94,208
outputs

(xp + xn) The performance of the technique is improved when the

Accuracy � , PSNR level is high and BER is low. For instance, LeNet and
(xp + yp + xn + yn)
AlexNet have low PSNR values (5.77 and 7.08), whereas
Recall xp ResNet and DenseNet achieved nearly 16 of PSNR. But the
� , proposed ThimuNet achieved 17.74 of PSNR. The existing
Sensitivity (xp + yn)
(12) techniques such as LeNet and AlexNet achieved high BER
xp (i.e., nearly 28.5), where ResNet and DenseNet achieved 11.9
Precision � , and 7.52 of BER. The proposed ThimuNet achieved very low
(xp + yp)
BER, i.e., 6.06, when compared with other existing tech-
Precision × Recall niques. The reason is that the proposed ThimuNet is well
F1 − Measure � 2 , adapted with all input features. Table 3 and Figure 12 show
Precision + Recall
the experimental results of proposed ThimuNet on preci-
where xp is denoted as True positive, xn is denoted as true sion, recall, sensitivity, specificity, and F-measure.
negative, yp is described as false positive, and yn is rep- In the precision experiments, the LeNet achieved 63%,
resented as false negative. AlexNet achieved 84%, ResNet achieved 91%, and DenseNet
The reason for comparing various pretrained models on achieved 94%, but the proposed ThimuNet achieved only
CNN is that they are significantly more accurate than the 92%. The reason is that some features are wrongly classified,
custom-built model of CNN. In addition, the pretrained and some collected cursive writings are not properly re-
models can effectively train on large datasets and directly use moved by using preprocessing techniques. After that, recall
the weights and architecture obtained to detect the cursive and sensitivity experiments are carried out to validate the
characters of Tamil on palm leaves. LeNet is the base for all performance of the proposed ThimuNet. The existing
ConvNets, which is mainly used for detecting the handwritten technique LeNet achieved 65%, ResNet achieved 79%,
characters. ResNet is used to solve the vanishing gradient AlexNet achieved 47%, and DenseNet achieved 86%,
problem by making the CNN to construct with more than whereas the proposed ThimuNet achieved 89% recall and
thousands of convolution layer without the increase of sensitivity. This proves that the proposed net achieved better
training error percentage and outperform the shallower performance than existing techniques. Finally, specificity
networks. AlexNet is developed by using eight layers with and F-measure experiments are performed. All the existing
learnable parameters, which has the ability to leverage GPU techniques achieved nearly 85% to 90% of F-measure and
for training and being able to train with vast numbers of 93% to 99% of specificity, where the proposed ThimuNet
parameters. In DenseNet, each layer is connected to every achieved 90.51% of F-measure and 99% of specificity. Table 4
other layer, and it also alleviates the vanishing gradient and Figure 13 provide the comparative analysis of proposed
problem. Moreover, the features are reused in this model, ThimuNet with existing techniques in terms of accuracy.
feature propagation is strengthened and minimized the The above table and Figure clearly prove that the pro-
number of parameters. Initially, the proposed CNN perfor- posed ThimuNet achieved higher performance than existing
mance is validated with the existing technique in terms of techniques in terms of accuracy. For example, the LeNet
PSNR and BER [23], tabulated in Table 2 and Figure 11. achieved 71% accuracy, ResNet achieved 88% accuracy,
12 Computational Intelligence and Neuroscience

Table 2: Validated results of proposed model with existing techniques.

Network type PSNR BER
LeNet 5.77 28.42
ResNet 15.69 11.9
AlexNet 7.08 29.74
DenseNet 16.08 7.52
ThimuNet 17.74 6.06

35

30

25

20

15

10

0
LeNet ResNet AlexNet DenseNet ThimuNet
Methodology
PSNR
BER
Figure 11: Graphical representation of proposed ThimuNet in terms of PSNR and BER.

Table 3: Experimental results of proposed ThimuNet with existing techniques.

Network type Precision (%) Recall (%) Sensitivity (%) Specificity (%) F-measure (%)
LeNet 63 65 65 79 63.87
ResNet 91 79 79 99 85.01
AlexNet 84 47 47 93 57.45
DenseNet 94 85 86 99 90.14
ThimuNet 92 89 89 99 90.51

AlexNet achieved 66% accuracy, DenseNet achieved 92% 0.95 seconds for the classification. However, the proposed
accuracy, and the proposed ThimuNet achieved 94% ac- ThimuNet achieved less prediction time (0.80 seconds) than
curacy. Finally, the prediction time of proposed ThimuNet existing techniques. The proposed ThimuNet achieved better
for classifying the cursive Tamil Palm Leaves Manuscript is performance from all experimental analysis and classified
given in Table 5 and Figure 14. wrong cursive letters due to irrelevant features. This problem
The existing LeNet, AlexNet, and ResNet prediction time can be solved by incorporating efficient feature selection
achieved nearly 1.08 seconds, while the DenseNet achieved techniques with proposed ThimuNet.
Computational Intelligence and Neuroscience 13

120

100

80

60

40

20

0
LeNet ResNet AlexNet DenseNet ThimuNet

Precision Specificity
Recall F-measure
Sensitivity
Figure 12: Performance analysis of proposed ThimuNet with existing techniques.

Table 4: Comparative analysis of proposed ThimuNet on the basis of accuracy.

Network type Accuracy (%)
LeNet 71
ResNet 88
AlexNet 66
DenseNet 92
ThimuNet 94

100

80

60

40

20

0
LeNet ResNet AlexNet DenseNet ThimuNet
Methodology

Accuracy
Figure 13: Graphical representation of proposed ThimuNet in terms of accuracy.
14 Computational Intelligence and Neuroscience

Table 5: Prediction time of proposed ThimuNet for classification of cursive Tamil palm leaves manuscript.
Network type Prediction time (s)
LeNet 1.06
ResNet 1.09
AlexNet 1.04
DenseNet 0.95
ThimuNet 0.80

1.2

0.8
Prediction Time (s)

0.6

0.4

0.2

0
LeNet ResNet AlexNet DenseNet ThimuNet
Methodology
LeNet DenseNet
ResNet ThimuNet
AlexNet
Figure 14: Comparative analysis of proposed ThimuNet in terms of prediction time.

5. Conclusion approaches have been used in each stage of the recognition

process.
Tamil is the world’s oldest languages and has attainment This work is one of the first attempts to build cursive
popularity among scholars throughout the world due to its training datasets by manual and automated segregation of
historical relevance and capacity to persist for centuries. Tamil palm leaf scripts. Future researchers could use the data
Individual interests and ideologies of academics have shaped set to develop expert systems that could then be used for a
study on palm leaves manuscripts in Tamil literature over range of purposes, such as a character classification based on
the periods and in a variety of ways. The most prevalent topic the century they have evolved and the identification of
of research in the actual world is cursive character recog- extinguished characters, and the identification of characters
nition of ancient Tamil characteristics, and it focuses on the whose shape could have changed. The research attempted to
documentation of Tamil characters to guarantee that a large segregate ancient Tamil palm leaf scripts to produce a
amount of data is gathered. However, due to improper care, massive amount of Tamil cursive character information.
palm leaf manuscripts have been destroyed, as numerous There have been no previous investigations on the identi-
people held them by default. As a result, a rigorous ex- fication of cursive characters in Tamil palm leaf scripts. The
amination of Tamil palm leaf texts, translation, and cata- ThimuNet of CNN is used in this research project to classify
loguing of their publishing were necessary. Thousands of cursive Tamil characters, and its performance is confirmed.
palm leaf manuscripts exist in the hands of institutes and In order to improve the performance of proposed model, the
separate healers, but they must be digitized, and proper weight of the input parameters must be optimized, which is
catalogues are developed for the future. Similarly, numerous considered as future scope of this research model. In ad-
researches on the cursive handwriting recognition of the dition, preprocessing plays a major role in removing un-
Tamil language are included in the survey. There is, however, wanted cursive writings; therefore, an effective
no universal technique for identifying cursive Tamil letters preprocessing technique is implemented along with the
(with sufficient accuracy) in their whole. As a result, various developed model for better performance.
Computational Intelligence and Neuroscience 15

Data Availability Sree Sankaracharya University of Sanskrit, Kalady,” Journal

Impact Factor, vol. 5, no. 1389, pp. 90–99, 2016.
The data used to support the findings of this study are [15] R. S. Sabeenian, M. E. Paramasivam, P. M. Dinesh, R. Adarsh,
available from the corresponding author upon request. and G. R. Kumar, “Classification of handwritten Tamil
characters in palm leaf manuscripts using SVM based smart
zoning strategies,” in Proceedings of the 2nd International
Conflicts of Interest Conference on Biomedical Signal and Image Processing,
pp. 18–21, Kitakyushu, Japan, August 2017.
The authors declare that they have no conflicts of interest. [16] B. Kiruba, A. Nivethitha, and M. Vimaladevi, “Segmentation
of handwritten Tamil character from palm script using his-
togram approach,” International Journal of Informative and
References Futuristic Research, vol. 4, no. 5, pp. 6418–6424, 2017.
[1] D. Ganapathy, “Preserving India’s palm Leaf Manuscripts for [17] S. Ghosh, A. Mahajan, and S. Banerjee, “Palm leaf manuscript
the Future,” WAGLOBAL, Kerala, India, 2016. conservation, the process of seasoning with special reference
[2] N. S. Panyam, V. L. T.R., R. Krishnan, and K. R. N.V., to Saraswati Mahal library, Tamilnadu in India: some tech-
“Modeling of palm leaf character recognition system using niques,” International Journal of Information Movement,
transform based techniques,” Pattern Recognition Letters, vol. 2, pp. 122–128, 2017.
vol. 84, pp. 29–34, 2016. [18] N. P. Challa and R. V. K. Mehta, “Applications of image
[3] K. P. Geena and G. Raju, “View-based feature extraction and processing techniques on palm-leaf manuscripts—a survey,”
classification approach to Malayalam palm leaf document Helix: The Scientific Explorer, vol. 7, no. 5, pp. 2013–2017,
image,” International Journal of Innovative Research in 2017.
Computer and Communication Engineering, vol. 2, no. 5, [19] R. Vinoth, R. Rajesh, and P. Yoganandhan, “Intelligence
pp. 264–267, 2014. system for Tamil Vattezhuttuoptical character recognition,”
[4] R. Chamchong and C. C. Fung, “A framework for the se- International Journal of Computer Science Engineering and
lection of binarization techniques on palm leaf manuscripts Technology, vol. 8, no. 4, pp. 22–26, 2017.
using support vector machine,” Advances in Decision Sciences, [20] M. Sornam and M. D. Poornima, “Tamil palm-leaf manuscript
vol. 2015, Article ID 925935, 7 pages, 2015. character segmentation using GLCM feature extraction,” In-
[5] N. P. Challa and R. V. K. Mehta, “Applications of image ternational Journal of Computer Science and Engineering, vol. 6,
processing techniques on palm leaf manuscripts-A survey,” in no. 4, pp. 167–172, 2018.
Proceedings of the Conference on Cognitive Science and Ar- [21] R. S. Sabeenian, M. E. Paramasivam, R. Anand, and
tificial Intelligence, CA, USA, February 2017. P. M. Dinesh, “Palm-leaf manuscript character recognition and
[6] S. Athisayamani, A. R. Singh, and A. S. Kumar, “Recurrent classification using convolutional neural networks,” in Com-
neural network-based character recognition system for Tamil puting and Network Sustainability, pp. 397–404, Springer,
palm leaf manuscript using stroke zoning,” in Inventive Singapore, 2019.
Communication and Computational Technologies, pp. 165– [22] M. Khosravy, N. Gupta, N. Marina, I. K. Sethi, and
176, Springer, Singapore, 2021. M. R. Asharif, “Morphological filters: an inspiration from
[7] R. S. Ratheash and M. M. Sathik, “A detailed survey of text line natural geometrical erosion and dilation,” in Nature-inspired
segmentation methods in handwritten historical documents Computing and Optimization, pp. 349–379, Springer, Cham,
and palm leaf manuscripts,” International Journal Of Com- Switzerland, 2017.
puter Sciences And Engineering, vol. 7, p. 99, 2019. [23] D. K. Thind and S. Jindal, “A semi blind DWT-SVD video
[8] J. Ali and J. T. Joseph, “A convolution neural network-based watermarking,” Procedia Computer Science, vol. 46,
approach for recognizing Malayalam handwritten charac- pp. 1661–1667, 2015.
ters,” Malayalam Handwritten character recognition using
cnn, vol. 9, no. 12, 2018.
[9] P. K. S. Balakrishnan and L. Pavithira, “Multi-font optical
character recognition using Deep Learning,” International
Journal of Recent Technology and Engineering, vol. 8, 2019.
[10] M. A. Hossain and S. Afrin, “Optical character recognition
based on template matching,” Global Journal of Computer
Science and Technology: C Software & Data Engineering,
vol. 19, 2019.
[11] K. Baskar, “Manuscript Libraries in Tamil Nadu: a study of
their organisation and preservation in the digital environ-
ment,” Doctoral Dissertation in Information Science, Madras
University, Tamilnadu, 2018.
[12] S. Devika and K. Vijayakumar, “Digitization of palm leaf
manuscripts in Tamil Nadu (India): a study,” Journal of Li-
brary Science and Research (JLSR), vol. 2, no. 1, pp. 1–10, 2016.
[13] R. Narenthiran and P. Ravichandran, “Cataloguing and
digitization of multilingual manuscript libraries in Tamil
Nadu: an evaluative study,” Journal of Advances in Library
and Information Sciences, vol. 5, no. 3, pp. 248–253, 2016.
[14] T. K. M. Sageer and A. T. Francis, “Analysis of the palm leaves
manuscripts collection for digital archiving: a case study of