Handwritten Bangla character recognition using convolutional neural networks: a comparative study and new lightweight model

Handwriting is a crucial way to enhance character recognition and learn new words. However, the Bangla characters consist of very complex shapes and similar patterns. Deep learning (DL) techniques have become a prominent solution for handwritten Bangla character recognition (HBCR) due to their ability to extract high-level features from complex data. Several DL techniques have been proposed for HBCR, but they are computationally expensive and large in model size and thus not suitable for use in resource-constrained devices such as smartphones. In this study, we have evaluated the state-of-the-art DL models for HBCR. For this, we have used four existing datasets and created a merged dataset (by combining the four) for cross-dataset evaluation. We have provided a comparative performance analysis of the state-of-the-art DL models for HBCR. Additionally, we have proposed a new lightweight DL model for HBCR and evaluated its performance. The proposed DL model consists of 74 layers, including sub-layers, and its architecture is divided into five similar blocks. It includes the convolutional layers of (3, 3) and (5, 5) kernels, (1,1) stride, and the maximum pool layer of the (2, 2) pool size. The proposed model achieved accuracy, model size, loading and testing times of 96.87%, 13 MB, 9.11 s, and 7.95 s, respectively. The experimental results show that our model outperformed state-of-the-art models in terms of efficiency (loading and testing time) and model size with competitive accuracy.