OngkoNet: An Ensemble Learning Approach for Recognizing Bangla Handwritten Digits

Handwritten Digit Recognition (HDR) plays a pivotal role in various fields, from digitizing historical documents to enabling automated data processing in machine learning applications. Researchers continue to develop diverse HDR pipelines by integrating various image processing techniques with conventional architectures based on machine learning (ML) and deep learning (DL). Traditional methods often require meticulous feature engineering and lack scalability for large and diverse datasets. Furthermore, the structural complexities of Bengali digits may result in ambiguous HDR scenarios. This paper presents an ensemble learning approach, named OngkoNet, to enhance Bangla handwritten digit recognition across multiple datasets. Unlike previous approches that achieve high performance on a single dataset, OngkoNet demonstrates consistent and superior performance across diverse datasets. The OngkoNet architecture incorporates three distinct models: ResNet50 (Retrained), Mo-bileNet (Retrained), and a custom Convolutional Neural Network (CNN) model. Through weighted average ensemble techniques, the strengths of each model are harnessed to enhance the overall recognition performance. To further improve evaluation robustness, we curated the Ongko dataset from diverse sources. Comparative experiments on NumtaDB, Ekush, BanglaLekha, and Ongko datasets confirm that OngkoNet significantly outperforms individual models. These results establish OngkoNet as a highly generalizable and efficient solution for Bangla HDR, paving the way for future advancements in multilingual HDR systems.