Topology-Aware Segmentation Methodology for Multi-Touched Handwritten Kannada Characters
TL;DR
This study presents a novel dual-algorithmic method, specifically designed for Kannada characters based on their boundary topology, which significantly outperforms the pre-trained YOLOv11 and the Watershed algorithm, which achieved success rates of 52% and 20%, respectively.
Abstract
Accurate character segmentation is significant for handwritten Kannada Optical Character Recognition (OCR) due to complex conjunct consonants and touching letterforms. This study presents a novel dual-algorithmic method, specifically designed for Kannada characters based on their boundary topology. Characters are classified into the Upper-Closed Curve (UCC) and Upper-Open Curve (UOC) categories. For UCC characters, the methodology identifies optimal cut spots using the highest blue peak vertical traversal criterion combined with Vertical Projection Profile (VPP) and thinning analysis of the top 30% segment. For UOC characters, where VPP is unreliable, a color-stuffing technique is employed for structural analysis. The dual-algorithmic was evaluated using a dataset of 900 challenging handwritten words containing touching characters from both topological categories. The results were compared against the traditional Watershed algorithm and a pre-trained You Only Look Once, version 11 (YOLOv11) model. The YOLOv11 model was trained over 100 epochs using a custom 18K-image dataset representing 62-character classes. With a 90% segmentation success rate, the proposed topology-aware algorithmic approach significantly outperforms the pre-trained YOLOv11 and the Watershed algorithm, which achieved success rates of 52% and 20%, respectively. Furthermore, the implementation of a hybrid pipeline, using the proposed topology-aware algorithmic approach, segmented multi-touching characters before YOLOv11 recognition, markedly boosts the performance of the deep-learning model. By resolving the major issues previously encountered in recognizing multi-touched components, this specialized front-end eliminates segmentation ambiguities that typically degrade accuracy in end-to-end systems. Consequently, this framework offers an effective solution for the challenge of recognizing multi-touch characters in complex scripts like handwritten Kannada.