A Cost-Effective Approach to Bengali Handwritten Character Recognition with Custom Convolutional Neural Network

Handwritten character identification is a crucial area in pattern recognition and image processing, given its broad range of applications. Although extensive research has been conducted on English character recognition, there is a noticeable gap in studies focusing on Bengali characters. This paper proposes a systematic method for recognizing Bengali characters. Our approach leverages a convolutional neural net-work (CNN) architecture designed specifically to manage the intricate structure of the Bengali language. Unlike previous studies, which often rely on larger datasets, our approach centers on focused experimentation using a curated dataset of 3,000 character samples. The recognition technique encompasses segmentation, feature extraction, and classification, emphasizing the latest advancements in CNN-based text detection for English characters. Notably, our research extends these techniques to Bengali handwritten characters, demonstrating the versatility and efficacy of our approach. Experimental results from our cu-rated dataset and CMATERdb 3.1.2, carefully divided for testing and training, were promising. We show that our model effectively identifies handwritten Bengali characters, as evidenced by high accuracy rates of 95.67% and 93.83%, respectively, with only 0.8 million parameters based on the number of output classes. The aforementioned results underscore the methodology's potential for both practical applications and future research projects.