A Deep Convolutional Neural Network Approach to Recognize Bangla Handwritten Digits


The demand for sorted recognition of manually written digits is increasing as a result of digitization in standard of living. Manually typed digit confirmation is essential for a few operations in several sectors. There was a great demand for a large and fair-minded dataset, which is why Bengali Handwritten Digit Acknowledgment may be an undervalued field. But in this paper, We have used the large and unbiased dataset of Bengali digits that is Numta DB. The challenging part of Numta DB is that it is unprocessed and highly augmented. In this paper, We have surfed through some of the techniques to pre-process the data and have trained it in our custom designed Deep CNN model. We have achieved a accuracy of 99.6% on the training set and 98.65% on the validation set. We have also created a test dataset of our own consisting of 289 samples. We achieved a accuracy of 87.20% on our own dataset.

Keywords multilabel; ensemble; incorporating multiple clustering centers; gated recurrent neural networks; temporal convolutional neural networks; long short-term memory.

[Full Paper]