Abstract:
Hand composed Digit recognition in Bangla language is a valuable beginning stage for
creating an Optical Character Recognition in the Bengali language. Be that as it may, absence
of huge and honest data collection, recognition of Bangla digit was not build already. In any
case, in this outline, a colossal & honest data source known as NumtaDB is utilized for
recognition of Bengali digits. The troublesome endeavor is connected to getting the solid
presentation and high precision for gigantic, fair, common, natural and particularly extended
NumtaDB dataset. So various sorts of preprocessing frameworks are utilized for planning
pictures and a significant convolutional neural network is utilized for the request of
representation in this paper. The LeNet-5 architecture based convolutional neural network
model has indicated superb execution. We have accomplished 97.5% testing exactness which
is a decent outcome for huge and fair NumtaDB dataset contrasting with other one-sided
datasets. A wide range of preprocessing of pictures is additionally significant before
preparing. We utilize some preprocessing strategies for obscure and loud pictures yet these
are insufficient for the elite. An examination of the system brings out the EMNIST and
MNIST datasets was performed so as to sustain the appraisal.