Research On Mongolian Online Handwriting Generation And Recognition Technology

With the rapid development of digital technology and the breakthroughs in pattern recognition in the field of machine vision,applications such as the recognition of various language and the identification of handwriting have emerged in recent years.However,due to the lack of reliable datasets,the research process of online handwriting recognition in Mongolian has been particularly slow.Based on the word formation of Mongolian,the vocabulary of Mongolian can reach millions,which makes it difficult to construct an online handwriting dataset containing the entire Mongolian words.To deal with these problems,this thesis study on the generation of Mongolian online handwriting samples.Due to the lack of publicly available large Mongolian online handwriting datasets and the prerequisite for generation experiments is a reliable dataset,this thesis constructs and make public a Mongolian online handwriting dataset(MOLHW).And to validate the dataset and achieve the generation of online handwriting samples,the studies based on MOLHW are as follows:(1)To ensure the reliability of the MOLHW dataset,this thesis validates the dataset for the handwriting recognition task and proposes an Encoder-Decoder combined with Attention-based online handwriting recognition model for Mongolian as the baseline model for the recognition task of the dataset.In addition to the current mainstream recognition models,handwriting recognition is also investigated on this dataset using a Transformer-based recognition model and a Long Short-Term Memory(LSTM)-based recognition model combined with a Connectionist Temporal Classification(CTC).The experimental results show that the dataset achieves good recognition results in all the above recognition models.(2)Aiming at the problem of online Mongolian sample generation,this thesis uses LSTM combined with Mixture Density Networks(MDN)model to achieve the generation of Mongolian online handwriting samples.The generated sequences are combined with the original dataset as the new training set,and the handwriting recognition experiments are conducted on the optimal baseline model,and the recognition result of the same test set is significantly improved.The experimental results show that the model is able to achieve the task of automatic generation of Mongolian handwriting samples for a given label.However,the model suffers from the problem of unstable generation quality and still requires manual work for sample selection.(3)Aiming at the problem of LSTM combined with MDN generation models,an online Mongolian handwriting sample generation model based on Generative Adversarial Networks(GAN)is proposed in this thesis.The model adds a word recognition module to the original GAN model structure.The loss of the generated samples through the word recognition module is introduced on the original loss of the generator.Experimental results show that the model can effectively achieve the task of sample generation for in-set words on a small sample set with high generation quality,which offers the possibility of scaling up to large datasets.