A THESIS SUBMITTED TO THE SCHOOL OF GRADUATE STUDIES IN PARTIAL FULFILLMENT OF THE REQUIREMENT FOR THE AWARD OF THE DEGREE OF MASTER OF SCIENCE IN COMPUTER SCIENCE

Abstract

Amharic is an indigenous Ethiopic script that follows a unique syllabic writing system adopted from an ancient Geez script. The Ethiopic script used by Amharic has about 317 different symbols of which 238 basic characters, 50 labialized, 20 numeric, and 9 punctuation marks. Recently Optical Character Recognition for the Amharic Script has become an area of research interest because there is a bulk of handwritten Amharic documents available in libraries, information centers, museums, and offices. The digitization of these manuscripts allows existing language technologies to be used to local information demands and advances. Limited research works have been made for handwriting character recognition of Amharic scripts but most of them use a dataset that is composed of text characters only, not including digit and punctuation mark scripts. A fully handwriting character dataset for Amharic scripts which include all text, digit, and punctuation marks is not available. As a result, doing complete handwriting character recognition at this level is very challenging and time consuming. So, this research will concern on handwriting digit and punctuation mark scripts only. In this research work, we develop a model for recognizing digit and punctuation mark scripts so that future researchers can integrate this research with previously done handwriting text character recognition and generate the complete handwriting character recognition for the Amharic language. For this research work, 200(two hundred) different handwritten Amharic digit and punctuation marks for each character (20 numerals and 9 punctuation marks) which is a total of 5800 (200x29) were collected. We used data augmentation technique to increase the training data. Using a convolutional neural network and by performing a grid search optimization on the hyper-parameters of the network, the researcher attained an accuracy of 96% for training, 95% for validation and tasting dataset