Segmentation-Free Multi-Font Printed Manchu Word Recognition Using Deep Convolutional Features and Data Augmentation

摘要

Precise Manchu character segmentation of segmentation-based Manchu recognition methods is difficult to realize because of complex Manchu language spelling rules and existence multi Manchu fonts. To avoid the influence of incorrect segmentation, this work proposes the idea of segmentation-free recognition to recognize Manchu word instead of Manchu characters. In addition, an end-to-end 9-layer convolutional neural network is proposed to automatically extract deep hierarchy features on Manchu word image. The proposed recognition model is applied on Manchu words with 12 Manchu fonts to evaluate its ability of multi-font recognition. Deep neural network needs massive data for training, whereas Manchu language is an endangered language lacking in document data. To solve this contradiction, this work firstly builds a Manchu dataset prototype and a multi-font Manchu word testing set, and then designs a data augmentation system to generate synthetic data for training. The data augmentation system contains 7 generation methods, including character structure distortion and image quality transformation. Experiments demonstrate the proposed convolutional neural network for Manchu word recognition achieves a new state-of-the-art accuracy on multi-font printed Manchu word. For printed Manchu fonts, the highest recognition accuracy reaches 0.95; the lowest accuracy is 0.88; the average accuracy of printed Manchu fonts reaches 0.91. Experiments also demonstrate the proposed data augmentation system is an effective way to solve insufficient data problem.