Using Deep Learning to Recognize Handwritten Thai Noi Characters in Ancient Palm Leaf Manuscripts

摘要

Extracting knowledge from ancient palm leaf manuscripts is essential for historians and other scholars who would like to access accumulated knowledge in the Thai Noi language manuscripts. In the absence of Thai Noi language readers, computer technologies play an important role in fulfilling this need. This research aims to apply deep learning approaches to recognize Thai Noi characters written in palm leaf manuscripts. The experiments were carried out by firstly collecting the page images of the manuscripts archived in the Museum of Art and Culture of Loei. Then the page images were preprocessed by converting to grayscale. To recognize Thai Noi characters, four convolutional neural network models based on inception and mobilenet networks namely Inception-v3, Inception-v4, MobileNetV1, and MobileNetV2 were evaluated. Handwritten Thai Noi characters were segmented from the grayscale images based on 26 Thai Noi characters. In this process, 100 images of each character were segmented and the whole dataset contained 2,600 images. Two image augmentation methods were applied to increase the amount of training data. Three experiments were carried out with three different datasets based on a 10-fold cross-validation design. The results indicate that MobileNetV1 outperformed other models in all experiments with an accuracy rate higher than 90%, while MobileNetV2 showed an interesting performance, which was almost equivalent to MobileNetV1 in the last experiment.