Classification of Dhamma Esan Characters By Transfer Learning of a Deep Neural Network

Abstract

We present an image classification of Dhamma Esan characters by fine-tuning the Inception V3 deep neural network trained on the ImageNet dataset. Dhamma Esan is a traditional alphabet used in the north-eastern region of Thailand, primarily written on Corypha leaves for the purpose of recording Buddhist scriptures. Preservation of these historical documents calls for the ability to classify the characters of the alphabet in order to facilitate digital indexing and searching, as well as assist anyone trying to read them. Our dataset consists of over 70,000 Dhamma Esan character images, much larger than any previous work. The result of ten-fold cross-validation showed that our model had 100% accuracy for four folds, and 99.99% for the other six folds. The previous best accuracy reported was 97.77%. We also developed a Dhamma Esan character classification web service where users can upload images of characters and get immediate classification results as well as mapping to the modern Thai alphabet.