Improving Sinhala OCR using Deep Learning

Abstract

Converting a printed document into a stream of characters using optical character recognition techniques is a widely researched area. However, the unique cursive property of Sinhala characters makes character recognition a challenge. An improvement in Sinhala OCR is a muchneeded requirement since this could benefit applications such as digitizing printed documents and books, data entry for business documents and, assistive technology for blind and visually impaired users among others. Even though accuracies of over 80% has been reported in previous studies, these have considered only a subset of Sinhala characters. In, this study we consider all the Sinhala characters including complex characters. Since languages such as English and other, Latin-based languages have achieved state-of-the-art accuracies in character recognition using deep learning, their application to improve the Sinhala optical character recognition of printed characters has not been explored. A contourbased segmentation method is used in this research to segment Sinhala characters and its output recognized using a Convolutional Neural Network. An overall accuracy of 85.37% was achieved for segmenting and recognizing Sinhala characters. In summary, convolutional neural network-based model is capable of improving the Sinhala printed character recognition.