A deep learning approach to handwritten number recognition

Abstract

Nowadays, Deep Learning is one of the most popular techniques which is used in several fields like handwriting text recognition. This paper presents our propose for a handwritten digit sequences recognition system. Our system, based in two stage model, is composed by Convolutional Neural Networks and Recurrent Neural Networks. Moreover, it is trained using on-demand scheme to recognize numbers from digits of the MNIST dataset. We will see that, with these training samples is not necessary segment or normalize the input images. Average recognition results were on 88,6% of accuracy in numbers of variable-length, between 1 and 10 digits. This accuracy is independent on the number length. Moreover, in most of the wrongly predicted numbers there was only one digit error.