Research On Large Vocabulary Offline Handwritten Mongolianholistic Recognition

In recent years,with the rapid development of digital technology,more and more Mongolian handwritten literature resources have been converted into image formats for preservation.However,Mongolian handwritten literature resources in image format is not suitable for word frequency statistics,content analysis,editing and retrieval.Therefore,this paper conducts related research on offline handwritten Mongolian recognition,and it can provide convenience for mining and utilization of Mongolian handwritten literature resources.The unique word formation of Mongolian language has led to a huge vocabulary.According to statistics,the Mongolian vocabulary size can reach several million,and the commonly used vocabulary is also hundreds of thousands.In addition,the offline handwritten Mongolian that this paper focuses on is unrestricted writing,and the word writing is very deformed,which makes the offline handwriting Mongolian recognition task extremely challenging.According to the characteristics of Mongolian vocabulary and serious writing deformation,this paper proposes a holistic recognition method suitable for Mongolian word formation and writing style.The main work of this paper is as follows:(1)The offline handwritten Mongolian holistic recognition is realized by using the most popular convolutional recurrent neural network and the connectionist temporal classification(Convolution Recurrent Neural Network-Connectionist Temporal Classification,CRNN-CTC).Tested on the relevant offline handwritten Mongolian dataset,and compared with the best performing deep neural network and hidden Markov model(Deep Neural Network-Hidden Markov Model,DNN-HMM)on the dataset,the experimental results show that the CRNN-CTC model is more suitable for the offline handwritten Mongolian holistic recognition task than the DNNHMM model.(2)According to the word formation and writing characteristics of Mongolian,this paper proposes a sequence-to-sequence model with attention mechanism and uses two different schemes for obtaining frames.The model regards the offline handwriting recognition task as the mapping problem of the handwritten word image frame sequence to the label sequence,and the attention mechanism can solve the problem of information loss caused by the input sequence being too long.Tested on the large vocabulary offline handwritten Mongolian dataset,the performance of the proposed model is not only better than the DNN-HMM model,but also better than the CRNNCTC model.The experimental results show that the sequence-to-sequence model with attention mechanism can not only overcome the influence of word writing deformation to a certain extent,but also solve the out-of-vocabulary problem effectively.Therefore,it is more suitable for completing the large vocabulary offline handwritten Mongolian holistic recognition task.(3)Aiming at the problem of insufficient sample of large vocabulary offline handwritten Mongolian dataset,this paper uses traditional spatial transformation method to achieve data augmentation.Using this data enhancement method to generate new samples,doubling the number of samples in the augmented dataset.Then,the data augmentation effect is verified on the sequence-to-sequence model with attention mechanism.The experimental results show that the data augmentation method proposed in this paper can further improve the recognition performance.