Handwritten Amharic characters Recognition Using CNN

• Published in: 2019 IEEE AFRICON pp. 1 - 4
• Main Author: Tamir, Kassahun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-09-2019
• Subjects: Amharic Characters; Character recognition; CNN; Convolutional neural networks; Handwriting recognition; handwritten recognition; Hidden Markov models; Testing; Training; Writing
• ISSN: 2153-0033
• DOI: 10.1109/AFRICON46755.2019.9133925

Amharic has been an official language of Ethiopia since many years ago. As a consequence, there are so many handwritten documents of several types, that are written in Amharic by hand, in monasteries, libraries, museums, universities, archives and with individuals. Among them, there are documents that have been written before several years and are getting older. To preserve these documents electronically the current way of digital preservation is not sufficient. Therefore, an intelligent handwritten classifier is essential. Furthermore, we can interact with latest smart devices such as with mobile phones by writing Amharic characters by hand by integrating this Amharic character classifier with them. Because contemporary smartphones are capable of reading and recognizing inputs written by hand on their modern input devices. Amharic alphabet has about 34 families of characters and about eight members with in a family, with some exceptions. There are, a total of, 286 different characters (286 classes). For this research, I have collected 30,446 characters from about 130 different individuals. Among them 27413 are used for training and the remaining 3033 are used for testing. Convolutional Neural Network algorithm with two convolutional layers is employed to classify the aforementioned characters to 286 classes. The design is evaluated on Keras and TensorFlow frameworks. After training the system, I have got a training accuracy around 99.52% after the first epoch, which is the least of accuracies of other epochs and a testing accuracy of 99.71%, approximately. In conclusion, the result of this research can be more improved by training the system using more data and using improved Convolutional Neural Network algorithms. In addition, the system is not evaluated using many different metrices such as by changing the orientation of characters and by adding noises; even though, they are written by different individuals and different styles.