Arabic Character Recognition Using CNN LeNet-5

Abstract

The human handwriting pattern is one of the research areas of pattern recognition; it is very complex. Therefore, research in this field has become quite popular. Moreover, human handwriting pattern recognition is needed for several things, one of them being character recognition. Recognition of Arabic handwriting is complex because everyone has different characteristics in writing and Arabic characters have quite abstract shapes and patterns. From previous research, Convolutional Neural Network (CNN), a deep learning-based algorithm, has a fairly high accuracy value when used for public datasets such as AHDB and private datasets. In this study, private datasets are used with a fairly high level of complexity because the respondents appointed to write Arabic letters come from different age categories. The CNN architecture used in this research is the architecture developed by Yan LeCun known as LeNet-5. The local dataset used was 8400 images, with details of 6720 for training data (each letter has 240 images) and 1680 for testing data (each letter has 60 images). The total respondents who wrote Arabic script were 30 people, and each person wrote each letter ten times. The accuracy obtained is 81% higher than in previous studies. The following study will test a number of additional CNN architectures to increase the accuracy of the results. In addition to accuracy, this study will also calculate the misclassification rate, root mean square error, and mean absolute error.