International Journal of Engineering and Manufacturing (IJEM)
IJEM Vol. 16, No. 2, 8 Apr. 2026
Cover page and Table of Contents: PDF (size: 2123KB)
Deep Learning-Based Recognition of Handwritten Devanagari Multilevel Compound Characters: A Performance Comparison
PDF (2123KB), PP.111-131
Views: 0 Downloads: 0
Author(s)
Index Terms
Handwritten Character Recognition, Devanagari Script, Optical Character Recognition (OCR), CNN, Attention-based CNN, LSTM, Compound Characters, Deep Learning, Machine Learning, Data Augmentation
Abstract
Handwritten character recognition is a crucial challenge in artificial intelligence and computer vision, particularly for complex scripts like Devanagari. Devanagari is widely used in Hindi, Marathi, and Nepali and consists of intricate multilevel compound characters, ligatures, and highly variable handwriting styles. Despite advances in optical character recognition (OCR) technology, accurately recognizing handwritten Devanagari characters remains difficult. This study compares various deep learning models, including Convolutional Neural Networks (CNN), CNN-SVM, Long Short-Term Memory (LSTM), EfficientNet, and a newly proposed attention-based CNN model. Extensive experiments were conducted on a diverse dataset containing simple and compound handwritten Devanagari characters. The proposed attention-based CNN model outperforms traditional methods, achieving a recognition accuracy of 96.5%, significantly higher than CNN (88.0%), CNN-SVM (88.5%), LSTM (92.0%), and EfficientNet (93.0%). The study employs advanced data augmentation techniques to enhance model robustness, making it adaptable to various handwriting styles. The attention mechanism in the proposed CNN model allows for improved feature extraction, leading to higher recognition accuracy and efficiency. This research contributes to developing robust OCR systems for the Devanagari script, enabling improved digitisation and preservation of Indian languages. The proposed approach can be extended to other complex scripts like Bengali and Tamil, further advancing multilingual OCR technologies, is now exploratory and has not been experimentally verified in this study. Future work can address the need for thorough cross-script evaluation and transfer learning studies to verify the adaptability of the attention-based CNN architecture, despite its inherent script-agnostic nature. The findings of this study hold significant implications for text digitization, historical document preservation, and automated language processing applications.
Cite This Paper
Ashwini B. Patil, Puneet Dwivedi, "Deep Learning-Based Recognition of Handwritten Devanagari Multilevel Compound Characters: A Performance Comparison", International Journal of Engineering and Manufacturing (IJEM), Vol.16, No.2, pp.111-131, 2026. DOI:10.5815/ijem.2026.02.07
Reference
[1]L. Schomaker, "Lifelong Learning for Text Retrieval and Recognition in Historical Handwritten Document Collections," Series in Machine Perception and Artificial Intelligence, WORLD SCIENTIFIC, pp. 221-248, Nov. 2020. doi: 10.1142/9789811203244_0012.
[2]V. Bansal and R. M. K. Sinha, "A complete OCR for printed Hindi text in Devanagari script," Proceedings of the Sixth International Conference on Document Analysis and Recognition, Seattle, WA, USA, 2001, pp. 800-804, doi: 10.1109/ICDAR.2001.953898.
[3]R. Thatte, "Handwritten Devanagari Character Recognition Using Deep Learning," rScroll, Oct. 24, 2022. doi: 10.47611/harp.228.
[4]D. Brezina, "Context-sensitive study of the visual similarity of characters in Cyrillic, Devanagari, and Latin scripts: effects of typeface design and expertise," Center for Open Science, Oct. 18, 2024. [Online]. Available: http://dx.doi.org/10.31234/osf.io/cfgdp.
[5]H. P. P. Win, P. T. T. Khine, and K. N. N. Tun, "Converting Myanmar printed document image into machine understandable text format," 2011 Sixth International Conference on Digital Information Management, Melbourne, VIC, Australia, Sep. 2011, pp. 96-101, doi: 10.1109/icdim.2011.6093371.
[6]A. Kulkarni and A. Shivananda, Deep Learning for NLP, in Natural Language Processing Recipes, Apress, 2021, pp. 213–262. doi: 10.1007/978-1-4842-7351-7_6.
[7]C. Bhole, "Devanagari Handwritten Marathi Compound Character Recognition System," Journal of Computer Based Parallel Programming, vol. 6, no. 3, pp. 1-8, Nov. 2021, doi: 10.46610/jocpp.2021.v06i03.001.
[8]A. Srivastav and N. Sahu, "Segmentation of Devanagari Handwritten Characters," International Journal of Computer Applications, vol. 142, no. 14, pp. 15-18, May 2016, doi: 10.5120/ijca2016909994.
[9]M. B. Sahin, D. M. Yalcinkaya, R. Dharmakumar, A. Hashemi, and B. Sharif, "Retrospective phase-map synthesis for CMR datasets from magnitude-only DICOM images enabled by A.I. generative models to create large training datasets for deep learning-based image reconstruction," J. Cardiovasc. Magn. Reson., vol. 26, p. 100987, 2024, doi: 10.1016/j.jocmr.2024.100987.
[10]S. S. Gaikwad, S. L. Nalbalwar, and A. B. Nandgaonkar, “Devanagari handwritten characters recognition using DCT, geometric and hue moments feature extraction techniques,” Sādhanā, vol. 47, no. 3, 2022. doi: 10.1007/s12046-022-01872-9.
[11]O. V. Astaf'eva, "A study into the governance specifics of educational institutions through the lens of the digitization effect on the educational process," National Interests: Priorities and Security, vol. 15, no. 4, pp. 772–783, Apr. 2019, doi: 10.24891/ni.15.4.772.
[12]K. Minatani, "The impact of presentation by voice media on reading speed when visually impaired persons read complex text including high-level content," Assistive Technology Research Series, vol. 34, pp. 659–664, 2013. doi: 10.3233/978-1-61499-304-9-659.
[13]Jubaer, S. M., Tabassum, N., Rahman, M. A., & Islam, M. K. (2023). Bn-drishti: Bangla document recognition through instance-level segmentation of handwritten text images. arXiv preprint, arXiv:2306.09351.
[14]Heil, R., & Nauwerck, M. (2023). Handwritten stenography recognition and the lion dataset. Springer Nature, arXiv:2308.07799v1.
[15]Retsinas, G., Sfikas, G., Gatos, B., & Nikou, C. (2024). Best practices for a handwritten text recognition system. arXiv preprint, arXiv:2404.11339.
[16]Shinde, A., & Dandawate, Y. (2020). Convolutional neural network-based handwritten Marathi text recognition. Journal of Xidian University, 14(8), 1585–1594.
[17]Agrawal, M., Chauhan, B., & Agrawal, T. (2022). Machine learning algorithms for handwritten Devanagari character recognition: A systematic review. Journal of Science and Technology, 7(1), 1–16.
[18]Sonkure, M., Gupta, R., & Moghe, A. (2020, October). An efficient approach for handwritten Devanagari script recognition. EasyChair Preprint, No. 4370.
[19]Gonwirat, S., & Surinta, O. (2022). Cycle augment: Efficient data augmentation strategy for handwritten text recognition in historical document images. Engineering and Applied Science Research. Retrieved from https://www.tcithaijo.org/index.php/easr/index.
[20]Saraswathi, D. D., & Sherif, S. M. (2021). Handwritten text recognition system using machine learning. Kristu Jayanti Journal of Computational Sciences, 1.
[21]Ehiorobo, E., Koleoso, R., & Uwadia, C. (2020). Training of offline handwritten text recognizers using computer-generated text. Elsevier Ltd.
[22]Khalkar, R., Dikhit, A. S., Goel, A., Gupta, M., & Patil, S. (2021). Deep learning for handwritten text recognition (ConvNet & RNN). Turkish Online Journal of Qualitative Inquiry (TOJQI), 12(8), 1874–1890.
[23]Sonawane, P., & Shelke, S. (2018). Handwritten Devanagari character classification using deep learning. 2018 International Conference on Information, Communication, Engineering and Technology (ICICET), 1–4.
[24]Aneja, N., & Aneja, S. (2019). Transfer learning using CNN for handwritten Devanagari character recognition. 2019 1st International Conference on Advances in Information Technology (ICAIT), Chikmagalur, India, 293–296.
[25]Bisht, M., & Gupta, R. (2020). Multiclass recognition of offline handwritten Devanagari characters using CNN. International Journal of Mathematical, Engineering and Management Sciences, 5(6), 1429–1439. Retrieved from https://doi.org/10.33889/IJMEMS.2020.5.6.106.
[26]Nagvanshi, S., Singh, A., Yadav, S., & Deepanshi. (2020). A detailed review of different handwriting recognition methods. International Research Journal of Engineering Technology, 7(3), 3462–3467.
[27]Deore, S., & Albert, P. (2020). Devanagari handwritten character recognition using fine-tuned deep convolutional neural network on trivial dataset. Sādhanā, 45, 243.
[28]P. G. et al. (2020). Convolutional neural network-based handwritten Devanagari character recognition. 2020 International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE). IEEE.
[29]Chakraborty, R., Mukherjee, D., Bhattacharyya, A., Mukherjee, H., Sur, M. K., Sen, S., & Roy, K. (2020). Online handwritten Bangla and Devanagari character recognition by using CNN: A deep learning concept. 68-71.
[30]Sawant, S., Sharma, A., Suvarna, G., Tanna, T., & Kulkarni, S. (2020). Word transcription of Modi script to Devanagari using deep neural network. 2020 3rd International Conference on Communication Systems, Computing and IT Applications (CSCITA 2020) - Proceedings.
[31]Dokare, I., Gadge, S., Kharde, K., Bhere, S., & Jadhav, R. (2021). Recognition of handwritten Devanagari character using convolutional neural network. 353-359.
[32]Saraswathi, D., & Sherif, S. M. (2021). Handwritten text recognition system using machine learning. Kristu Jayanti Journal of Computational Sciences (KJCS, 1(1), 58–69. Retrieved from https://doi.org/10.59176/kjcs.v1i1.2180.
[33]Pande, S., Jadhav, P., Joshi, R., Muddebihalkar, V., Rathod, S., Gurav, M., & Das, S. (2021). Digitization of handwritten Devanagari text using CNN transfer learning – A better customer service support. Neuroscience Informatics, 2, 100016.
[34]L. Li, "Convolutional Neural Networks (CNNs)-Based for Medical Image Analysis," in Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence, SCITEPRESS - Science and Technology Publications, 2024, pp. 546-552. doi: 10.5220/0012958600004508.
[35]P. Verma, R. Srivastava, and S. K. Tripathy, “An assessment towards 2D and 3D human pose estimation and its applications to activity recognition: A review,” SN Computer Science, vol. 6, no. 190, 2025. DOI: 10.1007/s42979-025-03708-9.
[36]U. Asgher, K. Khalil, Y. Ayaz, R. Ahmad, and M. J. Khan, "Classification of Mental Workload (MWL) using Support Vector Machines (SVM) and Convolutional Neural Networks (CNN)," in Proc. 2020 3rd International Conference on Computing, Mathematics and Engineering Technologies (iCoMET), Sindh, Pakistan, Jan. 2020, pp. 1-6, doi: 10.1109/icomet48670.2020.9073799.
[37]T. Joachims, A Statistical Learning Model of Text Classification for SVMs, in Learning to Classify Text Using Support Vector Machines. Springer US, 2002, pp. 45-74. doi: 10.1007/978-1-4615-0907-3_4.
[38]F. M. Salem, Gated RNN: The Long Short-Term Memory (LSTM) RNN, in Recurrent Neural Networks. Springer International Publishing, 2021, pp. 71–82. doi: 10.1007/978-3-030-89929-5_4.
[39]Kang, Rui & Park, Bosoon & Ouyang, Qin & Ren, Ni. (2021). Rapid identification of foodborne bacteria with hyperspectral microscopic imaging and artificial intelligent classification algorithms. Food Control. 130. 108379. 10.1016/j.foodcont.2021.108379.
[40]C.-Y. Zhu, Y.-K. Wang, H.-P. Chen, K.-L. Gao, C. Shu, J.-C. Wang, L.-F. Yan, Y.-G. Yang, F.-Y. Xie, and L. Jie, "A deep learning-based framework for diagnosing multiple skin diseases in a clinical environment," Frontiers in Medicine, vol. 8, Apr. 2021. doi: 10.3389/fmed.2021.626369.