Recognition of Handwritten Bangla Number Using Multi Layer Convolutional Neural Network
Md. Shovon, Md. Kamruzzaman, Mahantapas Kundu · 2020 IEEE Region 10 Symposium (TENSYMP) · 2020
In this paper, a new methodology is proposed to recognize the handwritten Bangla number using multi layer Convolutional Neural Network (CNN). The NumtaDB dataset is used for training the CNN model which consists of 85000+ digits images. After training the model with this vast, standard and unbiased data, the model attains a training accuracy of 99.66% and validation accuracy of 98.96% for recognizing bangla digits which shows that the model is not overfitted or underfitted and ready for real world unseen data. A great emphasis has been given on pre-processing and splitting of digits from a number since multi layer CNN alone is not good enough for attaining a significant accuracy. A small dataset containing 200 number images which has 2 digit to 5 digit numbers is tested against our proposed model. The validation accuracy for this model has been found to be 88%, which is higher than any recently proposed models.