A New State of Art Deep Learning Approach for Bangla Handwritten Digit Recognition using SVM Classifier

Deep Convolutional Neural Network (CNN) has earned remarkable success in computer vision technology, particularly in image classification task. With the growing interest, it has been also employed to recognize Bangla handwritten digits. In this paper, we have presented a deep convolutional neural network with support vector machine (CNN-SVM) in order to recognize Bangla handwritten digits. Here, we have designed and implemented a 14 layered deep CNN architecture as a feature extractor to acquire remarkable features automatically from Bangla handwritten digits image and SVM was used as a classifier to recognize Bangla handwritten digits. We have trained and tested the proposed model with two different Bangla handwritten digit dataset namely NumtaDB and BanglaLekha-Isolated. The accuracy was found 99.53% on NumtaDB dataset and 99.59% on BanglaLekha-Isolated dataset. We also used our prepared dataset which is completely unseen to proposed CNN-SVM model in order to evaluate the generalization capability of the proposed model. The CNN-SVM model achieved a test accuracy of around 99% on our prepared dataset that indicates that the proposed model is reliable and convincing.