Recognizing Bangla Numerals: A Deep Learning Approach on a Novel Handwritten Dataset

Optical character recognition (OCR) technology converts hand written or printed text into machine-readable digital text. Although OCR systems have advanced significantly in languages like English. Bangla requires enhanced recognition methodologies and superior datasets. This is mostly due to the distinct characteristics of Bengali numbers and the presence of compound numerals, which can be expressed in various forms. This research study aims to address the challenges and enhance the precision of Bangla handwritten numeral recognition by the application of a deep learning model. Models such as VGG16, VGG19, DenseNet121 and Ensemble technique have been employed for the recognition of handwritten numerals. The Primary dataset was compiled by our team and utilized to execute the model. We evaluated all 3 models on our primary dataset and a secondary dataset and attained exceptional performance, illustrating their efficacy and dependability and generalization in handwritten numeral identification. Among the evaluated individual models, VGG19 achieved the highest performance on primary dataset, attaining a test accuracy of 99.35%. VGG16 and DenseNet121 followed closely, each achieving a test accuracy of 99.19%. These outcomes underscore the robustness and effectiveness of deep learning architectures in the domain of handwritten numeral recognition. Furthermore, the implementation of an post-hoc ensemble technique, integrating the predictive capabilities of multiple models, resulted in an enhanced test accuracy of 99.52%, thereby affirming the significant potential of ensemble learning approaches in advancing the accuracy and reliability of handwritten character recognition systems.