Convolutional Neural Network Architecture for Offline Handwritten Characters Recognition

Abstract

Automatic handwriting recognition systems are a very wide-ranging research topic for many years. This type of intelligent systems is applied in various fields: Checks processing, forms processing, automatic processing of handwritten answers in an examination, etc. The purpose of this work is to propose a new Convolutional Neural Network (CNN) model and compare its performance with those of K-Nearest Neighbors (KNN) technique. Both used for the classification of complex multiclass problems. In this paper, we implement and evaluate the performance of this two-machine learning algorithms used to predict handwritten characters. The training and testing data have been extracted from the MNIST digits and letters database, which contains pre-processed images. The results obtained using different similarity measures such as accuracy, sensitivity and specificity confirm that the classification obtained by our proposed CNN architecture is the most accurate compared to the KNN studied in this work.