Deep learning for Ethiopian Ge'ez script optical character recognision

Ethiopie, or Geez is a Semitic language that has been spoken in the horn of Africa, in the history of Ethiopia, but now days it is only used as the language of liturgy in the Ethiopian and Eritrean Orthodox Tewahdo Church. Many of history, science and arts of Ethiopia have been documented by this language and almost all are found in the church. In this paper a Ge'ez Optical Character Recognition (OCR) using Convolutional Neural Network and a feed forward Multi-Layer perceptron are proposed. In order to apply these approaches two steps have done. Firstly, a Multi-Layer Perceptron neural network has developed to compare four widely used Ethiopic typefaces such as Ethiopian Jiret, Ebrima, Nyala and Abyssinica SIL. Images of 26 Ge'ez script written from these typefaces were given to the neural network for training and then scanned documents was given again as an input to check which of these typeface characters are recognized and classified. As a result, characters written in Ethiopian Jiret typeface has recognized very well. In the second part, using Ethiopian Jiret typeface many images of strings have been generated in using a standard python codec registry, then these images have used as an input to the Convolutional Neural Network and Multi-Layer Perceptron comparison. The results are promising and provide a better direction to the future work.

I. Introduction

Human beings have been faced different problems in areas like health, transportation, business and finance. Recently neural networks have been played a significant role in different application that can ease human beings' life such as medicine, natural language processing, business, computer vision and many other problems. Neural Network has grown from the very simple and most basic artificial neural network perceptron to the most modern and complex network with thousands and millions of neurons in order to simulate the ability of human brain to recognize and distinguish objects around. In fact, at this time many commercial scanners use Optical Character Recognition (OCR) systems to change an input image of character string to output a text file [4].