Bangla Hand Written Numeral Digit Recognition Using Deep

Abstract

As per the present scenario of Bangladesh, it can be stated that most Multi-National Companies (MNCs) or other sectors primarily use analogue etiquettes or hand-written forms. Moreover, the information being inscribed in those forms has to be re-input by a person inside the digital machine for further purpose. On the other hand, the existing hand-written models are using the erstwhile deep-learning algorithm which can hardly be labeled as compatible to the newer requirements. Although they serve the generic necessities regarding primal digit recognition, they are unlikely to provide the companies with significant accuracy in complex situations. This supposedly occurs due the lack of versatile training dataset in this kind of software. This paper investigates the prospects of an alternative solution namely the Convolution Neutral Network (CNN) as the proposed hypothesis. This probable mechanism may ensure an accuracy more than 96% compared to the former counterparts, particularly in regards with the most challenging and noisy cases. Using deep neutral network like the propounded CNN solution may fuse the gaps vis-á-vis the question of efficiency on the digital platform. From all around Bangladesh, approximately 72000+ specimens for training dataset have been collected along with 1700+ for the test dataset which can be depicted to have more accuracy than all other existing forms of solutions. Among the 14000+ specimens being used, there are ample amounts of noisy dataset included as well which can ensure the paramount accuracy in the least expected contexts. In this paper, a comparative analysis will also be presented explaining how the proposed model does have the best precision tactics in contrary with the other options. The objective is to provide a worthwhile courseware for the industrial, marketing and other interacting platform where quantitative and qualitative information require a distinct degree of accuracy.