Research On The Recognition Of Multilingual Ancient Characters In Natural Scenes Based On Deep Learning

The ancient pictogram records the culture and history of China,and is an important research object in the field of linguistics,ethnology and philology research.For the study of ancient pictogram,manual research is traditionally used,which is inefficient and is not conducive to recording,communication and spread.It is becoming more and more urgent to use advanced digital technology to study ancient pictogram.The research topic of this paper comes from the "2014 National Social Science Fund Project"(14ZDB104).The goal is to achieve image-based text recognition and retrieval,and build an overall database.It is convenient for users to use and communicate,and improve the efficiency of research.The project involves 22 languages(the least about 81 symbols and the most more than 2,000 symbols),for the most part does not included in the "Chinese Ancient Textual Catalogue" published in 1990(China Social Sciences Press,Chinese Ancient Chinese Character Research).For the study of ancient pictogram recognition,I have conducted a lot of research and comparison in the previous period.However,no relevant information has been found in automatic recognition of ancient pictogram.Finally,I selected the "deep learning" technology which is excellent performance in many visual tasks to complete the research.This paper adopts the idea of incremental learning and explores a complete set of engineering practice plan,which is divided into three stages:The first stage,a comparison was made between the deep learning technology and the traditional methods,and realized the task of 10 classification recognition based on convolutional neural network.Then,summarized the application of deep learning technology in the study of ethnic pictogram.In the second stage,applied the transfer learning technology to the field of ancient pictogram recognition,and solved the problem of fewer samples and more classifications of Ancient Pictogram Recognition.The third stage,study the popular Generative Adversarial Networks in the past one or two years.The paper used a Generative Adversarial Networks based on convolutional neural network.Randomly generated samples to optimize the model,making the model capable of handling pictogram in natural scenes.After three stages of research,a good test result has been achieved on the test dataset.Our work mainly includes:1.According to our research topic,a solution to the problem of small sample,multi-Classification and natural scene recognition is proposed.2.For the insufficiency of training data,according to the different requirements,three kinds of data enhancement schemes are proposed.3.Combined with engineering practice,the structure and parameters of convolution neural network are studied in depth.The network evaluation methods,such as visualization,sample feature space and network expressiveness,are summarized,and the network optimization scheme is arranged.Finally,the paper realizes the retrieval system of ancient pictogram based on the Web,and fully considers the extensibility of the system.It has lay a foundation for the follow-up work.