2024 International Conference on Distributed Computing and Optimization Techniques (ICDCOT)

Handwritten Telugu Character Recognition using

Machine Learning
1st Bhavesh Madhusudhan Karapu 2nd Anoop G L
Department of Computer Science & Engineering Department of Computer Science & Engineering
Christ Deemed to be University Christ Deemed to be University
Bengaluru, India Bengaluru, India
[email protected] [email protected]
2024 International Conference on Distributed Computing and Optimization Techniques (ICDCOT) | 979-8-3503-8295-2/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICDCOT61034.2024.10515646

3rd Manu Elappila 4th Mithun B N

Department of Computer Science & Engineering Department of Computer Science & Engineering
Christ Deemed to be University Christ Deemed to be University
Bengaluru, India Bengaluru, India
[email protected] [email protected]

Abstract—The Telugu language is the most prominent data requires less physical space and financial resources than
representative within the Dravidian language family, maintaining physical files. It reduces costs and eliminates the
predominantly spoken in the southeastern regions of India. need to organize, classify, and locate information in paper
Handwritten character recognition in Telugu has significant documents.
applications across diverse fields such as healthcare,
administration, education, and paleography. Despite its Character recognition applications include e-pdfs and
importance, the Telugu script differs significantly from English, signature verification; the insurance and banking industries
presenting distinct challenges in recognizing characters due to digitize forms, tax receipts, and transaction histories; the retail
its complexity and diverse character shapes. This study explores sector saves bills and customer transaction data. Healthcare
the application of machine learning, particularly delving into institutions utilize digital data tactics, such as electronic health
deep learning techniques, to improve the accuracy of Telugu records (EHR), to reduce errors brought on by unreadable
character recognition. This paper proposes a model to recognize prescriptions. Logistics organizations use HWR technologies
handwritten Telugu characters using Convolutional Neural to sort packages by detecting tags on shipments and scanning
Network (CNN). The proposed study demonstrates the accuracy bills of lading.
in identifying diverse handwritten Telugu characters. We assess
the system's performance against conventional and machine II. LITERATURE SURVEY
learning methodologies and preprocess an extensive dataset to
guarantee strong model training. The proposed model excels in This section gives a detailed survey of the research carried
accurately predicting visually similar but distinct characters, out in handwritten character recognition using different
achieving an impressive accuracy rate of 96.96%. Machine Learning (ML) techniques to recognize English and
various Indian scripts.
Keywords—Handwritten Telugu character recognition, Konkimalla et al. [1] used CNN to recognize the character
Convolutional Neural Network (CNN), character recognition.
in Telugu. The dataset used by Pramod et al. [2] contains 1000
I. INTRODUCTION words but could only cover some of the words in Telugu. For
the dataset of the characters, they have used the work done by
In recent years, machine learning has grown at a rapid Achanta and Hastie. However, they have yet to consider all
pace. One area that has gained more attention is Handwritten the combinations of status and guninthas. It scans the letter in
Character Recognition (HCR). Of all the languages spoken all segments and forms the network to recognize it. The author
over the world, Telugu, a Dravidian language spoken mainly has used two CNN architectures for classifying. Initially,
in the Indian states of Andhra Pradesh and Telangana, offers identify the character and further organize the vattu and
special potential and problems for HCR researchers. Telugu gunintham. Initially, in pre-processing, author performed
presents formidable obstacles to precise handwritten character skew correction, that correct the orientation or tilt image to
identification due to its complex script and variety of character straight, followed by segmentation and classification. Here are
shapes. With the development of machine learning some algorithms that the author uses: straight line though
algorithms, especially deep learning approaches, scientists transform for skew correction, a modified version of Otsu’s
have investigated creative ways to address these issues and version of binarization (removal of noise), for word
improve the accuracy of Telugu character recognition segmentation they are using Maximally Stable External
systems. Region (MSER) with some modifications, and for character
The proposed study explores machine learning approaches segmentation they are using Connected Components
for handwritten character recognition of Telugu characters, Algorithm.
including a thorough examination of current methods, Sharmila et al. [3] used OCR (Optical Character
drawbacks, and prospective directions for further Recognition) to recognize Telugu characters. Character
investigation. recognition takes in the six phases of scanning and digitizing,
The benefits of character recognition include Improved pre-processing, segmentation, feature extraction,
data storage. It makes room for the finest data storage classification, and post-processing [2]. It scans and saves the
available. Handwritten data, such as original signatures or data in a standard form for later stages. In preprocessing, it
notes, can be electronically translated into other forms of data enhances the document image, preparing it for later stages to
in many papers, contracts, and personal records. Electronic achieve higher accuracy. In this, it removes the tilting of the

979-8-3503-8295-2/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.
image by using skewing techniques and removes the noise networks to achieve state-of-the-art error rates. Sonthi et al.
present in the image. Segmentation determines the parts of the [10], the researchers employed a Multi-Objective Mayfly
characters in the image. Many OCR algorithms segment the Optimization with Deep Learning (MOMFO-DL) approach
words into isolated characters. Feature extraction measures for character recognition from given images. Specifically,
the attributes of patterns that are pertinent to the given Sonthi et al. [10] focuses on Telugu character recognition
classification, and in classification, it is the decision-making using MOMFO-DL. The model incorporates DenseNet-169 as
stage of an OCR; it takes the features extracted from the a feature extractor, generating a set of informative vectors.
previous stage and compares them with the text of preset rules Additionally, a Functional Link-Neural Network (FLNN)
and in post processing the output is converted to ASCII or serves as the classification model for recognizing and
ISCII or other coding schemes so that these are reconstructed categorizing printed characters. The optimization technique
from outputs of classification stage. A well-structured MOMFO is applied to fine-tune the model's parameters,
dictionary can also be used to resolve the ambiguities in aiming to enhance overall performance in the task of Telugu
recognition. character recognition. This hybrid approach combines deep
learning with multi-objective optimization to achieve more
Muppalaneni [4] used the CNN technique for handwritten effective and accurate character recognition results. CCS
character recognition. Here, invariant features are produced by
concepts are used in this model (computer methodologies ->
the convolution of the image and the filters and then passed on computer graphics, shape modeling).
to the following layers. The features in the subsequent layer
are twisted using various filters to produce more invariant Ganji et. al., [11] used VGGNET(Visual Graphic) for
features, and this process is repeated until the final feature is recognizing Ancient Telugu script name Golusu Kattu which
obtained. The convolution layers suggest that every part of the is from kakatiyan time. Ganji et al. [11] recognize the golusu
picture might have an intriguing pattern since regions are just kattu writing (ancient Telugu script in Kakatiya times).
continuous blocks of pixels. Another essential aspect is deep- Several documents consisting of ancient knowledge, health
CN adaptability and ability to function well with image input. care tips, wealth information, and several land records can be
retrieved using this technique. This uses a method based on
Burra et al. [5] presented a technique in two ways to deep learning. Algorithms like transfer learning are used.
improve the symbol or glyph segmentation in a Telugu OCR. VGGNET(Visual Graphic group-16) can be trained and can
Segmenting is the most essential part of a Telugu OCR be used as a deep feature extractor. CNN is used to extract
system; its impact is on the OCR system's performance. A high-level features, but sufficient is required. Kesana et al.,
glyph is a character's specific shape, design, or representation. [12] a HOG feature and Bayesian classification are used to
In initial stage segmented word into valid glyphs using identify individual handwritten Telugu characters. Using this
conventional ML, two-class SVM (support vector machine). algorithm, the author got a recognition rate of 87.5% from
The second hypothesis is supported by the physical Telugu scripts. Bayesian networks can produce higher
positioning of the interconnected components. The Burra et al. recognition rates using nearest Neighbourhood methods.
[5] have tested over 5000 pages over 30 books. Even though Nagarajan et al. [13] utilized the TCR model (DLTCR-
some of their books could have been of better quality and had PHWC) for character recognition, capable of identifying both
many ill glyphs, their results showed significant printed and handwritten characters. The recognition process
improvements. Burra et al. [5] uses the zone information to begins with image pre-processing employing the adaptive
segregate connected components intern merge the broken fuzzy filtering technique. Feature extraction is carried out
glyphs. through a fusion of EffectiveNet and CapsuleNet models. The
Mahesh [7] used a cloud-based optical character recognition stage involves the use of the Aquila optimizer
recognition pattern based on early technique and AIML. The (AO) in conjunction with a bidirectional long short-term
main purpose of [6] is to help people persistently type data memory (BiLSTM) model.
from meaningful images and extract text from it. The author Gupta et al. [14] have used the segmentation method to
[6] uses A Telegram bot to transfer picture-based electronic identify cursive alphabets. First, the alphabets are primarily
records or Images based on text. In the backend, the data is segmented into a particular character and then detected and
extracted into organized, writable information. This can be
merged to form a meaningful phrase by comparing it with a
useful for people in the data entry field. The above process thesaurus. They used only 26 phrases for their work, and
uses Google Tesseract, which is basically a tool used for hence, it is limited to those only.
optical character recognition for many Languages to extract
text from the given picture dataset. Singh et al., [8] Artificial Bozinovic et al. [15] have utilized a holistic technique,
Neural networks are used by dividing the given Input Image which signifies phrases via various transformation phases like
into separate lines, words, and characters. This paper deals letters, features, contours, phrases, and points. Partially
with the OCR for printed Telugu scripts. The system takes an calculated characters are identified through lexicons
image as input, and then it separates the lines, words, and containing only 130 words. Hence, it is restricted to these
characters in a stepwise manner. Then, it recognizes the words only. Arica et al. [16] applied the HMM technique to
characters using an artificial neural network approach. identify the characters and utilized the hybrid technique to
make the best use of HMM. Each and all alphabets are
Achanta et al. [9] have used deep convolutional neural inspected in four different directions to extract features.
networks that classify the characters and extract lines from an Pradeep et al. [17] employ Neural Networks for character
image. In this work, an end-end Framework segments the recognition, utilizing diverse techniques including
characters using mathematical morphology, classifies the backpropagation neural network, nearest neighbor network,
characters, and extracts the lines. Here, the classification and radial basis function network. They compare the
model is a deep convolutional neural network, and the performances of these networks and augment the number of
language is modeled as the third-degree Markov chain at the neurons in the hidden layer to improve overall effectiveness.
glyph level. Achanta et al. [9] work on advanced neural

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.
 Patel et al. [18] employed a multi-resolution approach III. METHODOLOGY
utilizing Discrete Wavelet Transform (DWT) along with In the proposed method for recognizing the Telugu
Euclidean Distance Metric (EDM). The EDM calculates the handwritten characters, we first resize the input image so that
distance from each unidentified input vector to every mean all the datasets are the same size. Then, we convert the images
vector, and the smallest distance confirms the input vector's to black and white format. After the image processing part, we
cluster pattern. This method achieves an identification split the dataset into training and testing data. It also involves
precision of up to 90.77%. In the work by Michael et al. [19], segmentation and training of the model with the characters.
neural-based techniques are presented for alphabet Fig. 1. depicts the general flow of character recognition
recognition. Two neural models, incorporating directional and system.
transition features, are employed alongside two feature
extraction methods. The features are matched using Back
Propagation (BP) and Radial Basis Function (RBF) network
classifiers. For transition features, the feature vector
comprises 100 elements, while directional feature extraction
results in an 81-element vector. This approach is demonstrated
using the CAS dataset.
Hallale et al. [20] have compared traditional and
directional feature extraction techniques. They have utilized
12 directional features for identifying characters and digits.
Choudhary et al. [21] employ a method for identifying
lowercase English alphabets, utilizing binarized pixels of
images as features and employing multilayer backpropagation
neural networks as classifiers. The image undergoes filtering,
binarization, and resizing to a dimension of 15 x 12. Fig. 1. The general flow for Telugu Handwritten Character Recognition.
Consequently, a feature vector of size 180 is generated for
each character, serving as input for the neural network during The proposed method follows the pipeline below –
its learning process, and it achieves precision maximum character input- image resizing and normalizing- neural
85.62%. Cruz et al. [22] have recognized cursive alphabets. In network- and character recognition. Fig. 2. depicts the
this, the authors have extracted various characteristics. 9 methodology of the proposed work.
Characteristics are extracted. Every characteristic is A. Dataset
independently provided as input to nine multilayer networks,
Recognizing Telugu characters poses a significant
and outcomes are merged with every alternate many law.
challenge, primarily due to the requirement for extensive
The literature gap in the survey on Telugu character datasets to train Convolutional Neural Networks (CNN)
recognition lies in the comprehensive consideration of all effectively. We've developed a unique dataset encompassing
possible combinations of status and guninthas. While several 52 distinct Telugu letters to ensure the neural network is well-
studies have utilized convolutional neural networks (CNN) equipped for character recognition. The different Telugu
and optical character recognition (OCR) techniques for characters and the data collected sample are shown in Fig.3.
character recognition, they still need to address Telugu We have collected datasets from different age groups of
characters' nuances fully. Specifically, there needs to be more people, educated and uneducated people with different ink,
exploration of all combinations of status and guninthas, which styles, and fonts; a few samples are shown in Fig. 4.
are critical elements in Telugu script.

Fig. 2. The Proposed Architecture for Telugu Handwritten Character Recognition.

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.
 computational complexity. After the convolutional and max-
pooling layers, a flatten layer is introduced. This layer
transforms the 2D output from the preceding layers into a 1D
layer. The flattening step is crucial for transitioning from
convolutional layers to fully connected layers in the neural
network architecture.
The first dense layer takes the output of the flatten layer as
input and consists of 128 neurons. This layer is designed to
learn global patterns and representations based on the features
extracted by the convolutional layers. The final dense layer,
with as many neurons as there are classes in the classification
task, produces the ultimate predictions for the model. The final
Dense layer, often accompanied by a softmax activation
function, is responsible for producing an output probability
distribution over the different Telugu characters. Each neuron
in this layer corresponds to a specific character class. The
character with the highest probability is considered the
predicted output.

Fig. 3. 52-Telugu characters.

IV. RESULTS AND DISCUSSION
The result depicts the implementation of a model for
recognizing handwritten Telugu characters. Training included
all 52 characters of the Telugu language. We compiled a
dataset containing over 3000 images of Telugu characters
sourced from diverse age groups, fonts, colors, and styles for
experimentation.
After training the model for 30 epochs, the model achieved
a commendable test accuracy. The result of the confusion
matrix showed the model's performance for various Telugu
character classes. The matrix offers a thorough understanding
of the model's predictive power, which is shown in Fig. 5.
Accuracy is an important attribute to recognise a character.
Fig. 4. TeluguScriptPalette: A Kaleidoscope of Fonts, Colors, and Styles
Fig. 6. indicates the accuracy achieved through the training
B. Image resizing and Normalization process. The accuracy plot, which illustrates the training
process, shows how the model has improved across epochs
The Telugu character images undergo normalization and throughout the validation and training stages. To train the
resizing to a standardized size of 128 x 128. This resizing and proposed model, 80% of Telugu character images were
normalization process facilitates the neural network's training utilized, while the remaining 20% of images were reserved for
in a streamlined manner, ensuring simplicity and effectiveness testing purposes. The model successfully recognized Telugu
in the learning process. characters from the provided images, with each prediction
C. Convolutional Neural Network (CNN) systematically verified against the corresponding actual data
for accuracy. Fig. 7. depicts the examination of the model. The
Character recognition performance relies heavily on the proposed model demonstrates the ability to accurately predict
effectiveness of the network's training. Our network characters that share similar visual features but are distinct,
comprises several essential layers, including convolutional, such as "aa" and "pa," as depicted in Fig. 8. Despite the visual
max-pooling, flatten, and dense layers. similarities, the model can differentiate between these
The convolutional layers consist of two stages: the first characters correctly. Table 1 presents an overview of existing
layer incorporates 32 filters, each of size 3 x 3, and the second works and the proposed model, highlighting the methods
layer employs 64 filters, also of size 3 x 3. Both layers utilize employed, respective datasets utilized, and the accuracy
the Rectified Linear Unit (ReLU) activation function, achieved in character recognition. This analysis indicates the
introducing non-linearity to capture local patterns and features significance of the proposed method and dataset collection in
in the input Telugu characters. 32 filters slide over the image, handwritten character recognition. The equation (1) is used
detecting patterns such as edges, corners, and textures. Each evaluated proposed methodology.
filter produces a feature map, capturing different aspects of the
input. 64 filters layer capture more complex patterns and = (1)
relationships in the input data.
In each convolutional layer, a max-pooling layer with a Where:
pooling size of 2 x 2 is applied. This operation reduces the
TP: True Positives; TN: True Negatives; FP: False
spatial dimensions of the feature maps, aiding in extracting
Positives; FN: False Negatives.
essential Telugu character features while minimizing

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.
Fig. 5. The proposed model confusion Matrix

Fig. 7. (a) Input image. (b) Predicted outcome.

(a) (b)

Fig. 6. Accuracy plot for proposed method.

(c) (d)
Fig. 8. Recognizing the similar characters. (a) Input Telugu ‘aa’ character
image. (b) Recognized ‘aa’ character (c) Input Telugu ‘pa character image.
(d) Recognized ‘pa’ character.

(a) (b)

TABLE I. COMPARISON OF CHARACTER RECOGNITION METHODS AND ACCURACY RATES

Existing works Method Used Datasets Accuracy
Konkimalla et al. [1] CNN 32 images per category 96.7%
Naresh Babu [4] CNN Printed Telugu 52 characters 79.61%
Burra et al. [5] KNN Telugu printed fiction books 98.75%
Gupta et al. [14] Neural Network and SVM English cursive alphabets for 26 words 62.93%
Patel et al. [18] ANN cursive alphabets 98.46%
Hallale et al. [20] Neural Networks English cursive alphabets 85.62%
Proposed Work 52-handwritten Telugu characters 96.96%

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSION [10] Vijaya Krishna Sonthi, S. Nagarajan, and N. Krishnaraj, “An
Intelligent Telugu Handwritten Character Recognition Using Multi-
One of the complex and ongoing research topics in the Objective Mayfly Optimization with Deep Learning–Based DenseNet
fields of pattern recognition and image processing has been Model,” ACM Transactions on Asian and Low-Resource Language
Information Processing, vol. 22, no. 3, pp. 1–16, Mar. 2022.
handwriting recognition. Handling a dataset with several
[11] T. Ganji, M. S. Velpuru, and R. Dugyala, “Multi Variant Handwritten
classes is challenging. This study proposed a Telugu Telugu Character Recognition Using Transfer Learning,” IOP
handwritten character recognition method that follows a Conference Series: Materials Science and Engineering, vol. 1042, no.
structured pipeline encompassing character input, image 1, p. 012026, Jan. 2021,
resizing, normalization, and CNN for efficient training and [12] M. L. K, V. K, S. G, S. D, and D. P, “Hand Written Telugu Character
recognition. Developing a unique dataset comprising 52 Recognition Using Bayesian Classifier,” International Journal of
distinct Telugu letters ensures the neural network's Engineering and Technology, vol. 9, no. 3S, pp. 37–42, Jul. 2017.
proficiency in character recognition. The proposed method [13] V. K. Sonthi, S. Nagarajan, and N. Krishnaraj, “Automated Telugu
and its unique dataset emerge as a valuable contribution to Printed and Handwritten Character Recognition in Single Image using
Aquila Optimizer based Deep Learning Model,” International Journal
handwritten character recognition and can accurately predict of Advanced Computer Science and Applications, vol. 12, no. 12,
visually similar yet distinct characters. The proposed method 2021, doi: https://doi.org/10.14569/ijacsa.2021.0121275.
achieves accuracy as high as 96.96%. For future work, we can [14] Gupta, Anshul, Manisha Srivastava, and Chitralekha Mahanta. "Offline
collect more handwritten character images, which may help handwritten character recognition using neural network." In Computer
improve the accuracy of the proposed Telugu character Applications and Industrial Electronics (ICCAIE), 2011 IEEE
recognition system. A capsule network can also be included to International Conference on, pp. 102-107. IEEE, 2011.
handle photos in various orientations. [15] Bozinovic, Radmilo M., and Sargur N. Srihari. "Off-line cursive script
word recognition." PatternAnalysis and Machine Intelligence, IEEE
Transactions on 11, no. 1 (1989): 68-83.
REFERENCES
[16] ARICA, NAFIZ. "An off-line character recognition system for free
[1] C. Konkimalla, Y. Prakash, G. Srikar, S. Trishal, S. Mandal, and style handwriting." PhD diss., MIDDLE EAST TECHNICAL
Channappayya, “Optical Character Recognition (OCR) for Telugu: UNIVERSITY, 1998.
Database, Algorithm and Application.” Accessed: Apr. 28, 2023.
[17] Pradeep, J., E. Srinivasan, and S. Himavathi. "Neural network based
[2] Pramod Sankar K, CV Jawahar, and Raghavan Manmatha, “Nearest recognition system integrating feature extraction and classification for
neighbor based collection ocr,” inProceedings of the 9th IAPR English handwritten." International Journal of Engineering
International Workshop on Document Analysis Systems. ACM, 2010, Transactions B: Applications 25, no. 2 (2012): 99.
pp. 207–214.
[18] Patel, D. K., T. Som, and M. K. Singh. "Improving the Recognition of
[3] M. Sharmila, D. Assistant, and M. Gangadhar, “A comparative Study Handwritten Characters using Neural Network through Multiresolution
of Classification Algorithm for Printed Telugu Character Recognition,” Technique and Euclidean Distance Metric." International Journal of
International Journal of Electronics Communication and Computer Computer Applications 45, no. 6 (2012): 38-50.
Engineering, vol. 3, no. 3, Accessed: Apr. 28, 2023.
[19] Blumenstein, Michael, Brijesh Verma, and Hasan Basli. "A novel
[4] Naresh Babu Muppalaneni, “Handwritten Telugu Compound feature extraction technique for the recognition of segmented
Character Prediction using Convolutional Neural Network,” 2020 handwritten characters." In Document Analysis and Recognition, 2003.
International Conference on Emerging Trends in Information Proceedings. Seventh International Conference on, pp. 137-141. IEEE,
Technology and Engineering (ic-ETITE), 2020. 2003.
[5] Burra, Sukumar & Patel, Amit & Bhagvati, Chakravarthy & Negi, Atul. [20] Hallale, Sumedha B., and Geeta D. Salunke. "Twelve Directional
(2018). Improved Symbol Segmentation for TELUGU Optical Feature Extraction for Handwritten English Character Recognition."
Character Recognition. 10.1007/978-3-319-76348-4_48. International Journal of Recent Technology and Engineering 2, no. 2
[6] Kesana, Mohana & Babu, Tummala. (2019). A Novel Telugu Script (2013).
Recognition and Retrieval Approach Based on Hash Coded Hamming: [21] Choudhary, Amit, Rahul Rishi, and Savita Ahlawat. "Off-line
Proceedings of the International Conference on Communications and handwritten character recognition using features extracted from
Cyber Physical Engineering 2018. 10.1007/978-981-13-0212-1_58. binarization technique."AASRI Procedia 4 (2013): 306-312.
[7] Mahesh C. “Telugu Optical Character Recognition using Cloud [22] Cruz, Rafael MO, George DC Cavalcanti, and Tsang Ing Ren. "An
Computing And Python,” 2021. Accessed: Apr. 20, 2023. ensemble classifier for offline cursive character recognition using
[8] Singh, Rinki & Kaur, Mandeep. (2023). OCR for Telugu Script Using multiple feature extraction techniques." In Neural Networks (IJCNN),
Back-Propagation Based Classifier. The 2010 International Joint Conference on, pp. 1-8. Ieee, 2010.
[9] R. Achanta and T. Hastie, “TELUGU OCR FRAMEWORK USING
DEEP LEARNING.” Accessed: Apr. 20, 2023.

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on March 10,2025 at 13:25:55 UTC from IEEE Xplore. Restrictions apply.