Improving the Performance of Local Binary Convolutional Neural Networks for Bilingual Digit Recognition

Abstract

Bilingual handwritten digit recognition is a challenging problem that is not investigated deeply. In many applications, most public documents in Arabic countries are written in bilingual forms, Arabic and English. Arabic digit recognition and English digit recognition were studies using several machine learning techniques, including a convolutional neural network that is used in many applications and modified to produce other models such as Local Binary Convolutional Neural Networks (LBCNN). However, the parameters used for building the LBCNN model were not tuned in proportion to the LBCNN technique, including the number of depths or convolutional layers, anchor weights, sparsity level, and the learning rate approach. This paper enhances the performance, in terms of accuracy, of the LBCNN technique for both Arabic and English digit recognition models by introducing optimum parameters values. The proposed parameters enhance the accuracies of LBCNN models by 0.04% and 0.08% for MNIST and MADBase datasets, respectively.