Synthetic Data and DAG-SVM Classifier for Segmentation-Free Manchu Word Recognition

There are a few studies on Manchu recognition, and the existing methods are mainly based on segmentation on characters or strokes. Thus, their performances are strongly dependent on segmentation accuracy. In this paper, a whole word recognition method for segmentation-free Manchu word is proposed to avoid the mis-segmentation of Manchu word. Firstly, we build an initial Manchu word image dataset, and then augment it with synthetic data, which are harvested via structural distortions on Manchu word image. Secondly, the support vector machine classifier with polynomial kernel function combined with directed acyclic graph is used for classification of Manchu words from 2 to 100 classes. The experiment results show that the precise is 78% for the 100-way classification problem, even above 90% for classes less than 40. The synthetic data method proposed in this paper is an effective way to augment the training and test dataset for Manchu word recognition.