Thai Printed Character Recognition Using Long Short-Term Memory and Vertical Component Shifting

Abstract

The segmentation-based approach for Optical Character Recognition (OCR) works by first segmenting a text line image into individual character images and then recognizing the characters. The approach relies heavily on the performance of the segmentation process and thus suffers from the problem of touching and broken characters. On the other hand, the unsegmented approach for OCR processes the text line image without segmenting the image into individual characters, and the approach is more suitable for languages such as Thai that contains a lot of touching characters in nature. This paper proposes an application of Long Short-Term Memory (LSTM), which is an unsegmented method, to Thai OCR. The paper also introduces a method called vertical component shifting to solve the problem of a large number of vertically occurring character combinations that occur in four-level writing system of Thai, and pose difficulty for standard LSTM networks. The experimental results demonstrate the better accuracy of our proposed method over standard LSTM networks and other commercial software for Thai OCR.