Thai Handwritten Character Recognition Using Deep Convolutional Neural Network

A handwritten character recognition is a key for many applications that rely on digital documents, such as healthcare, insurance, and banking sectors. Deep learning is an innovative way of carrying out the handwritten character recognition that is typically performed by humans. This research discusses supervised deep learning using teachable machine, web-based tool, as a tool to train, evaluate and test a convolutional neural network (CNN) model for 44 Thai handwritten consonants. The best model is selected to apply in web and mobile application for practicing writing Thai consonants. A pre-trained model, Mobilenet, is used in the training. The Burapha-TH Thai handwriting dataset is used to train, evaluate and test the model. The experimental results show that the number of epochs and the number of images in dataset affects model accuracy. The dataset that has the maximum number of images, achieves the highest training accuracy, 99.53%, at 60 epochs and 0.001 learning rate and achieves the highest testing accuracy, 83.43%, at 70 epochs and 0.0001 learning rate.