BnBorno-MNet: A MobileNetV2 Architecture for Bangla Handwritten Character Recognition

Handwritten character recognition (HCR) of Bangla handwritten characters is very challenging due to the script's complexity. Many deep-learning methods have been used for this task, but they are often too large and slow for smart-phones. In this study, we present BnBorno-MNet, a lightweight MobileNetV2-based model designed to improve the recognition of basic handwritten Bangla characters and digits. A customized head of five layers is added to the base model, and transfer learning is applied by unfreezing the last 20 layers to fine-tune the network. The model is trained, validated, and tested using 11 vowels, 39 consonants, and 10 digits from the BanglaLekha dataset, and further evaluated in a zero-shot setting on the Ekush and CBD2023 datasets. Compared to the base MobileNetV2, BnBorno-MNet achieves significantly higher accuracy, improving by 13.29%, 25.21%, and 16.48% on BanglaLekha, Ekush, and CBD2023, respectively. These results demonstrate the robustness and strong generalization ability of the proposed model for Bangla handwritten character recognition.