Evil method: A deep CNN model for Bangla handwritten numeral classification

In recent times, Convolutional Neural Networks (CNN) have proven their ability to classify a complex dataset like that of handwritten numerals. Though CNN does not require any handcrafted feature extraction method to learn useful features, they may need a large number of sample images to train accurately. Typically, different augmentation methods are used to enlarge the training datasets in order to learn more useful features as well as to stop over training. The present work proposing a deep model and an image augmentation method for classifying handwritten Bangla numerals. Moreover, the proposed deep model was trained with an augmented and enlarge training set and then tested with a collection of 3996 image samples. The proposed model with the proposed augmentation method achieved a testing accuracy of 99.42% on its 74th iteration. In addition, the best weight from that trained model also used for cross validation on two different and independently collected Bangla benchmark datasets. The cross validation results on those particular data sets are 99.53% and 95.56% respectively.