MSBNet: Handwritten Bangla Character Recognition Using Lightweight Multi-scale CNN Architecture

With the rise in machine learning deployments in many real-world applications, the demand for handwritten character recognition is increasing rapidly. Most of the applications are currently restricted to the English language, but to broaden their impact, it is also essential to accommodate non-English speakers. This has led to an urgent need for character recognition systems for regional languages. The past few years have witnessed work in regional languages worldwide, ranging from Chinese and French to Indic languages. Within the field of Indic languages, the work on Bangla character recognition is in very nascent stages. The primary challenge is the enriched character set of more than 80 classes (including compound characters) and the variation in handwriting styles across different subjects. Most existing works are either limited to a subset of classes or have significant computational or storage overheads. To address this, the present paper proposes a novel lightweight multi-scale convolutional neural network (CNN), called \(\mathcal {MSB}\)Net, that generalizes well for a variety of real-world handwritten Bangla character datasets. Furthermore, \(\mathcal {MSB}\)Net can be smoothly integrated into edge devices due to its significant reduction in the number of parameters (correspondingly storage and computational requirements) compared to state-of-the-art (SOTA) architectures. The paper validates the efficacy of the proposed model experimentally and conducts an ablation study on the model.