Yorùbá Character Recognition System Using Convolutional Recurrent Neural Network

Abstract

Handwritten recognition systems enable automatic recognition of human handwritings, thereby increasing human-computer interaction. Despite enormous efforts in handwritten recognition, little progress has been made due to the variability of human handwriting, which presents numerous difficulties for machines to recognize. It was discovered that while tremendous progress has been made in handwritten recognition of English and Arabic languages, very little work has been done on Yorùbá handwritten characters. Those few works, in turn, made use of Hidden Markov Model (HMM), Support Vector Machine (SVM), Bayes theorem, and decision tree algorithms. To integrate and save one of Nigeria's indigenous languages from extinction, as well as to make Yorùbá documents accessible and available in the digital world, this research work was undertaken. The research presents a convolutional recurrent neural network (CRNN) for the recognition of Yorùbá handwritten characters. Data were collected from students of Kwara State University who were literate Yorùbá writers. The collected data were subjected to some level of preprocessing such as grayscale, binarization, and normalization in order to remove perturbations introduced during the digitization process. The convolutional recurrent neural network model was trained using the preprocessed images. The evaluation was conducted using the acquired Yorùbá characters, 87.5% of the acquired images were used for the training while 12.5% were used to evaluate the developed system. As there is currently no publicly available database of Yorùbá characters for validating Yorùbá recognition systems. The resulting recognition accuracy was 87.2% while the characters with under dot and diacritic signs has low recognition accuracy.