Classificação de caracteres manuscritos da base IRONOFF utilizando deep learning

Abstract

The handwritten character classification aims to recognize, from visual information, which character is being represented. This technique is used in important applications, such as checks processing in banks and zip code reading, which saves time for such repetitive tasks. The intrinsic characteristics of the handwritten process hampers this recognition, once manuscripts are different for each person, in terms of style, format and size. This research carried out a handwritten character classification using the characters present in the IRONOFF dataset, which has uppercase and lowercase characters. The experiment was made by using Deep Learning and by comparing different neural networks architectures. To do so, a convolutional neural network architecture was built, LeNet-5 was implemented and pre-processing operations were applied. The best results were achieved for the proposed network with the accuracy of 92,61%, 86,65% and 79,67%, for the subsets of uppercase, lowercase and both of them, respectively.