Handwritten Bangla Numeral Recognition Using Ensembling of Convolutional Neural Network

Despite being one of the major languages in the world, research regarding Bengali handwritten numeral recognition (BHNR) isn't enough in comparison with the other prominent languages. Existing methods mostly rely on feature extraction and some older machine learning algorithms. Recent bloom in machine learning due to deep neural network especially using Convolutional Neural Network (CNN) showing promising results in this field with better accuracy. Some recent works show very good accuracy only in recognizing plain simple digits but perform poor in challenging scenario because of lack of large and versatile training dataset. In this work, we've ensembled our best performing proposed CNN models to recognize numerals with high degree of accuracy beyond 96% even in most challenging noisy conditions. Initially 72000+ specimens from NumtaDB (85000+) have been used for training and 17000+ specimens have been used as test dataset. The improvement in performance in challenging scenarios has been observed, when various noisy training specimens have been augmented to create a training dataset of size about 114000 specimens. The performance of our proposed model has been compared with other existing works also and presented here. These finding are based on Computer Vision Challenge on Bengali HandWritten Digit Recognition (2018) competition submissions.