Smart Network-Based Classification of Handwritten Basic, Modified, and Complex Conjunct Characters in Devanagari Script
Year: 2025
Type: article
Abstract: The Devanagari script is extensively used for documenting information in both the government and private sectors in India. In Devanagari documents, modified and conjunct characters frequently appear alongside consonants. Intricate forms of these characters, combined with variations in writing patterns, make the recognition process a challenging task. Most Devanagari datasets available online lack modified and conjunct characters. To address this issue, a dataset comprising basic, modified, and complex conjunct characters of the Devanagari script, which includes 580 classes, is created. Additionally, a segmentation algorithm is developed to automatically segment filled forms from different writers, thereby accelerating dataset processing. To classify these characters accurately, three convolutional neural networks, SmartNet1, SmartNet2, and SmartNet3, were designed by experimenting with various hyperparameters, such as the number of layers, filters, nodes in the fully connected layer, and kernel size. Each SmartNet was modified using a support vector machine and a k-nearest neighbors classifier, yielding a total of nine configurations. All nine configurations were tested on both self-generated and benchmark datasets. These networks are relatively shallow and have fewer parameters, enabling faster convergence and accomplishing remarkable results across different datasets.
Authors: Pallavi Patil, Kevin Noronha
Institution: St. Francis Institute of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4417070682