A Deep CNN-Based Approach for Revolutionizing Bengali Handwritten Numeral Recognition

Recognition of Bengali handwritten digits is a fascinating and demanding research problem that has garnered significant interest from researchers in the fields of pattern recognition. In this paper, a task-oriented deep convolutional architecture for recognizing handwritten Bengali digits is proposed. The main goal is to get a high level of accuracy while using a small number of parameters. The proposed architecture is designed to address the challenges posed by the complex and diverse nature of handwritten numerals in Bengali script, while also being computationally efficient with only 1.08 million trainable parameters. The performance of the model was evaluated by conducting experiments on two commonly used benchmark datasets of handwritten numerals in Bengali script, CMATERdb-3.1.1 and BanglaLekha-isolated-numerals. Different augmentation techniques were utilized to enhance the diversity and size of the training set, which led to improved robustness and generalization of the model. On the CMATERdb-3.1.1 dataset, the proposed model achieved an accuracy of 99.28%, and on the BanglaLekha-isolated-numerals dataset, it achieved an accuracy of 99.12%, outperforming several state-of-the-art models with comparable or larger numbers of parameters. The results suggest that this task-oriented model can be an efficient and effective solution for the recognition of Bengali handwritten numerals, with potential applications in document analysis, digitization, and text recognition.