Improved Prototypical Network with Diacritic Detection for Few-Shot Urdu Handwritten Character Recognition
Seshendranath Balla Venkata, Kavitha Thiyagarajan, C K Sarumathiy., Saef Thallal, Mukesh Soni
Abstract
In recent years, few-shot learning has been confirmed for handwritten character recognition in low-resource languages by allowing classification with a minimal number of labelled samples. However, existing methods, such as the Enhanced Prototypical Network (EPN), often struggle with visually alike characters, particularly in Urdu scripts where small diacritic marks change meaning. Hence, this research proposes a Diacritic-Aware Prototypical Network (DA-PN) that integrates diacritic-aware weighting, targeted distance scaling, and lightweight post-processing detection to improve recognition accuracy. Initially, data were collected from the Urdu Printed and Handwritten Character Dataset (UPHCD), which consists of scanned images of Urdu characters written by various individuals in changing handwriting styles. The collected data were then pre-processed and encoded into a high-dimensional embedding space using a convolutional encoder. Subsequently, class prototypes were created using an enhanced weight distribution module that arranged samples with clear and correctly placed diacritics. Moreover, the distance-scaling approach with contrastive loss confirms that embeddings of similar but dissimilar characters are pushed farther apart, which improves the separation. Finally, a lightweight diacritic detection module confirms the predictions by ensuring the presence and position of the marks. The experimental results show that the proposed DA-PN framework attains an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 8. 3 \%}$</tex> for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1}$</tex>-shot when compared to the existing EPN model.