Recognition of Bengali Handwritten Digits Using Convolutional Neural Network Architectures

Handwritten digit recognition has been the "hello world" of deep learning. Yet, there are no significant work on Bengali handwritten digits due to a lack of benchmark dataset. NumtaDB is the largest dataset on Bengali handwritten digits and currently we have the best accuracy of 99.3359% on it. We used popular CNN architectures namely, ResNet34 and Resnet50. We preprocessed the data, used data augmentation, and trained our models with augmented data. We tested our models on both the raw test data and cleaned test data. We found that slightly underfitted models work better on the test data. And finally ensembled our six best models to get our final predictions. In this paper we describe some methods and techniques that performs well in NumtaDB dataset.