A Topological Feature-Based Segmentation for Multi-Touched Handwritten Indic Words: A Case Study on Kannada Scripts

Abstract

The challenge associated with developing a reliable Optical Character Recognition (OCR) system for the Indic languages, such as Kannada, lies in the complex morphological composition that features round shapes and complex conjuncts called ‘Ottaksharas’. Although deep neural networks are effective in identifying objects from a broad perspective, they typically lack topological awareness, which is necessary to segment overlapping characters at the pixel level. At present, overlapped / multi-touched units are generally perceived as opaque concepts by traditional end-to-end frameworks without any consideration of their linguistic and geographical context. This research attempts to overcome these limitations with an innovative topology-aware dual algorithmic approach based on “upper-closed” and “upper-open” curve geometries. Using this approach, the algorithms for the Feature Preserving Bisection (FPB) and Adaptive Ratio Based Ottakshara Segmenter (ARBOS) determine the optimal paths for segmenting characters. For recognition purposes, a manually annotated dataset of 18,600 characters belonging to 62 distinct categories was compiled to train a YOLOv11 network, obtaining 91.0% segmentation effectiveness in difficult multi-touch scenarios. The proposed hybrid solution produced a final recognition accuracy of 92.0%, providing a 12.0% increase compared to the baseline scenario and a character error rate (CER) of 8.0%. This finding is supported by the empirical analysis of the Jaccard Index, which reveals the superior performance of the framework against U-Net and Mask R-CNN by 7.3% and 6.1%, respectively.