Synergistic Hybrid Segmentation for Handwritten Kannada Word Recognition Addressing Deep Learning Challenges

Title

Synergistic Hybrid Segmentation for Handwritten Kannada Word Recognition Addressing Deep Learning Challenges

Creator

Malini, M.; Hemanth, K.S.

Description

The handwritten Kannada script has an intricate aksharas that are formed by combining consonants, vowels, and ottus. These complex combinations pose significant hurdles for automated text segmentation. The inherent diversity in handwriting styles, coupled with prevalent character overlap, multi-touch connections, varied curve structures such as upper open curve(OC), upper closed curve (CC), and the highly condensed nature of Ottaksharas, routinely blurs character boundaries, leading to severe segmentation errors that propagate and compromise overall recognition accuracy. A hybrid approach that customizes adaptive traditional methods like vertical pixel count, to identify true character gaps in handwritten Kannada characters could effectively manage character overlap, or segment multi-touch characters or Ottaksharas. This pre-processing stage can allow subsequent deep learning models to recognize this segmented character. This will avoid significant hurdles: immense data requirements for pixel-level annotations, high computational costs for dense prediction, and significant architectural complexities for precise boundary delineation and handling connectivity. Given these constraints, particularly with less-resource language like Kannada, scaling deep learning models will lead to ever erroneous recognition. This paper argues that modified traditional approaches, by directly embedding customized knowledge and leveraging targeted feature engineering, can offer a computationally efficient and data-lean alternative. This strategy enables more robust segmentation for complex Kannada characters, providing a practical pathway for automated handwritten text processing in such linguistic domains. 2025 IEEE.

Source

2025 5th Asian Conference on Innovation in Technology, ASIANCON 2025;

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

character overlap; deep learning challenges; handwritten text segmentation; hybrid approaches; Kannada script; Ottakshara; traditional image processing

Coverage

Malini M., Reva University, Computer Science and Applications, Rukmini Knowledge Park, Karnataka, Bangalore, India; Hemanth K.S., Christ University, Department of Computer Science, Karnataka, Bangalore, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-835035699-1;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/ASIANCON66527.2025.11281074

https://www.scopus.com/pages/publications/105031419257?origin=resultslist