Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition

EasyChair Preprint 15539

Abstract

Few-shot learning (FSL) enables effective classification with minimal examples per class, offering a valuable solution for languages with limited annotated datasets. Traditional research has predominantly focused on improving performance with deep learning models on large-scale datasets. However, creating extensive datasets for all languages is often impractical and time-consuming. Few-shot learning approaches can provide good results with minimal data. This research focuses on Bangla and Hindi characters and numerals due to their complex and intricate structures aiming to demonstrate FSL's applicability to other languages with scarce handwritten datasets. Given the complexity of Bangla and Hindi characters, this study uses these languages to test the effectiveness of FSL models, under the premise that a model adept at handling Bangla's complexity will also perform well with other languages of similar or lesser complexity. This paper introduces SynergiProtoNet, an advanced hybrid network designed to enhance the recognition of handwritten characters and digits. We extensively benchmark various state-of-the-art few-shot learning models, including BD-CSPN, Prototypical Network, Relation Network, Matching Network, and SimpleShot. SynergiProtoNet integrates advanced clustering techniques and a robust embedding mechanism to capture intricate details and contextual nuances. Specifically, it utilizes multi-level (high level & low level) feature extraction within a prototypical learning framework. Our comprehensive analysis demonstrates that SynergiProtoNet significantly outperforms all other few shot learning models across various experiments—Monolingual Intra-Dataset Evaluation, Monolingual Inter-Dataset Evaluation, Cross-Lingual Transfer, and Split Digit Testing a new benchmark in the field.

Keyphrases: Handwritten Character Recognition, Prototypical Network, SynergiProtoNet, cross-lingual transfer, few-shot classification

@booklet{EasyChair:15539,
  author    = {Mehedi Ahamed and Radib Bin Kabir and Tawsif Tashwar Dipto and Mueeze Al Mushabbir and Sabbir Ahmed and Md.Hasanul Kabir},
  title     = {Performance Analysis of Few-Shot Learning Approaches for Bangla Handwritten Character and Digit Recognition},
  howpublished = {EasyChair Preprint 15539},
  year      = {EasyChair, 2024}}