Online Handwritten Mongolian Word Recognition Using a Novel Sliding Window Method with Recurrent Neural Networks

Abstract

Because of the conglutinated characteristic of Mongolian words, it's difficult to realize online handwritten Mongolian word recognition with high recognition accuracy based on segmentation-based strategy. Meanwhile, as the vocabulary of Mongolian words is large, using a segmentation-free method with deep bidirectional long short term memory(DBLSTM) network is more suitable. We design a 5 bidirectional hidden level DBLSTM network for online handwritten Mongolian word recognition. This paper mainly proposes a novel sliding window method which selects frames with different intervals to enhance recognition rate. The novel method can generate hundreds of sequence data for each sample, while only one sequence data is generated using ordinary sliding window method. More sequence data and more abundant sequence information are helpful to raise the recognition rate. We evaluated the recognition performance on our online handwritten Mongolian database with 925 classes. The proposed method achieves the word level recognition rate of 89.24% with PCA feature extractor and best path decoding, compared to that of 88.45% using ordinary sliding window method. Further, several well trained DBLSTM models based on the proposed method are combined to vote the output, finally, the word-level recognition raises to 90.35%.