Research On Deep Learning For Historical Mongolian Document Images Retrieval

Recently,with the development of multimedia,more and more historical Mongolian documents have been converted into digital images so that they can be preserved as long as possible.This not only beneficial to the preservation of historical Mongolian documents,but also provides a suitable way for historical researchers.However,the document images do not save the related index information,and can't realize the full text search of document which indirectly restricts the utilization of historical Mongolian documents.Therefore,this thesis will focus on the research of historical Mongolian documents images retrieval technology so as to improve the efficiency of historical Mongolian image retrieval.In recent years,Deep Learning(DL)has been widely used in the field of image retrieval.There are many classical deep learning network models being applied to feature extraction.The process of deep learning is based on the simulation of human brain mechanism to establish multi-layer neural networks and automatically extract data features.Convolutional Neural Network(CNN)is one of the classic models of deep learning.Because of its weight sharing characteristics,it greatly reduces the number of free parameters and improves the efficiency of learning features.Therefore,this thesis will take the Kanjur images as the object of study to explore the image retrieval technology of historical Mongolian document images.The specific research works are as follows:(1)Combined with the characteristics of Mongolian word formation,this thesis proposes an image normalization scale which is suitable for Mongolian word images representation.By comparing the four normalized sizes(28×28,32×32,32×64 and 32×96 respectively),when the historical Mongolian document image is normalized to 32×96,its retrieval performance is better than the other three normalizations.(2)Combined with the deep learning technology,this thesis proposes another new method which is called Convolutional Neural Network to express the features of historical Mongolian document images.Under the proposed CNN model,every Mongolian word image can be represented by the same length of feature vector.The experimental results show that the retrieval performance of proposed CNN is better than the LeNet-5,AlexNet,ZFNet,GoogLeNet and Auto Encoder method.It proves that the CNN model proposed in this thesis has certain advantages in the representation of historical Mongolian document images.