Applying a semi-supervised learning approach to reduce noise in Thai-OCR

Thai characters are one of the most complex characters because of many reasons. For example, they can be aligned into different levels, they are composed of a number of small components, and there are no word or sentence separating symbols. Noise reduction algorithms which are successfully applied to English documents might yield a poor result from Thai documents. This paper thus proposes a novel noise reduction method that is suitable for Thai documents using a semi-supervised learning approach. Results obtained from experiments shows that our method does not only obviously remove the noise but also preserve small components of Thai characters.