Research On Printed Mongolian Holistic Recognition Technology Based On Deep Learning

With the rapid development of computer technology,an increasing number of printed Mongolian documents have been scanned into image formats for conservation.However,document resources in image format cannot be further used for editing and retrieval.Optical character recognition(OCR)techniques can convert the Mongolian character into an editable text format.In the existing research on printed Mongolian OCR,it is common to adopt the segmentation-based recognition method.The Mongolian word to be recognized is segmented into characters,and then character segmentation is performed.Character segmentation is a difficult task.So,we can use end-to-end method for holistic recognition,thereby reduce the recognition errors caused by the segmentation errors.The main research contents are as follows:(1)The first part of this thesis is end-to-end based printed Mongolian holistic recognition.This study regards the Mongolian recognition problem as a sequence to sequence matching problem,in which input word image is treated as a multi-frame sequence and the output recognition result is regarded as a sequence of letters.This study introduces two deep learning models.One is the holistic recognition model that combines connectionist temporal classification with Long Short-Term Memory,and the other is sequence to sequence model that includes attention mechanism.The experimental results show that the proposed method is superior to the traditional segmentation-based recognition method and solve the issue of OOV.(2)The second part of this thesis is multi-task based holistic recognition and character segmentation.It is common for word recognition to determine the position of the character then identify it.It also implicitly performs this work of character segmentation in the holistic recognition.If this hidden process can be supervised,it would achieve better result.This study based on the sequence to sequence recognition model with attention mechanism requires the attention mechanism to complete the character segmentation task.The attention mechanism enhances the accuracy of character positioning and improves the overall recognition rate.The experimental results show that the proposed multi-task method can improve the accuracy of holistic recognition.Meanwhile,it is superior to the existing character segmentation method in the segmentation task and get better character segmentation results.