Research On Text Detection And Recognition Of Multi Fonts Printed Tibetan

The research on the detection and recognition of Tibetan text is one of the key directions of Tibetan informatization processing.Since Tibetan fonts can be roughly divided into two categories,namely,Wu Jin and Wu Mei fonts,most of the previous studies on the detection and recognition of Tibetan texts printed with multiple fonts firstly segmented Tibetan text images into Tibetan text line images horizontally using traditional processing methods to achieve the task of detecting Tibetan texts,then segmented Tibetan text line images vertically to obtain Tibetan Ding images,and finally achieved the task of Tibetan text recognition by classifying Tibetan Ding images.This research method has tedious experimental process,small scale of available training data resources as well as poor generalization ability and portability of experimental results,and mainly studies single fonts such as Tibetan Wu Jin.Due to the complexity of the Wu Mei fonts,the detection and recognition of the Wu Mei font is less studied.The classification of Wu Jin and Wu Mei of Tibetan fonts is only a big classification,They are subdivided into many different fonts,and the differences are still relatively large,plus many Tibetan text images contain multiple fonts,so it is of great theoretical and practical significance to carry out effective research on the detection and recognition of multi-font printed Tibetan.This paper lays the foundation for the work of this paper by analyzing the background and significance of the research on text detection and recognition of Tibetan multiple fonts,reviewing the current status of research on text detection and text recognition in Chinese and English,and focusing on the theoretical methods used in the research related to text detection and recognition of Tibetan.The main work of this paper is as follows.1.Collection and pre-processing of experimental data.A large amount of Tibetan corpus data is obtained by using crawling technology.By analyzing the encoding position of Tibetan characters in Unicode,the design program preprocesses the Tibetan corpus data to obtain corpus data containing only modern Tibetan syllables.At the same time,22 Wu Jin gold fonts and 15 Wu Mei fonts in Tibetan were collected.The text detection and recognition datasets of Tibetan in multi-font printing were constructed using statistical methods and image processing methods such as Open CV-Python.2.Detection of Tibetan in multi-font printing.The strategy of joint training with 0.5 sampling rate and DBNet pre-training model with ResNet-18 as backbone network in natural scenes are fine-tuned to complete the Tibetan text detection experiment.3.Recognition of Tibetan in Wu Jin multi-font printing.The three text recognition methods,CRNN,Rosetta and RARE,which are based on Mobile Net V3 and Res Net-34 backbone networks,are applied to the recognition of Tibetan in Wu Jin multi-font printing.Analyze the recognition performance of six comparative experiments on two test sets;On this basis,CRNN and SRN text recognition methods with Res Net-50 as the backbone network are applied to the recognition of Wu Mei multi-font printed Tibetan,and the recognition effects of two sets of comparative experiments on two test sets are analyzed.4.Design multi-font printed Tibetan recognition system.The recognition system is developed with the modular idea of "high cohesion and low coupling" software design,and an online recognition system with simple interface and easy to expand functions is realized by using open source technologies such as Python.The following research results were achieved in this paper:1.The text detection and recognition dataset for multi-font printed Tibetan language is constructed.Including 1200 Tibetan language detection datasets;2.88 million Wu Jin multi-font printed Tibetan recognition datasets and 2.88 million Wu Mei multi-font printed Tibetan recognition datasets.2.The detection of multi-font printed Tibetan is realized.The Tibetan detection dataset is applied in the DBNet pre-training model with Res Net-18 as the backbone network in the natural scene to realize the unified detection research on the images containing Tibetan Wu Jin and Wu Mei fonts.3.The recognition of multi-font printed Tibetan is realized.Proposed a strategy of using 74 Tibetan characters as a recognition transcription dictionary and an evaluation index for Tibetan text recognition;In the aspect of Wu Jin printing multi-font,the comprehensive comparison experiment results show that the overall effect of CRNN text recognition method under the same backbone network is better than Rosetta and RARE,and the CRNN text recognition method with Res Net34 as the backbone network is the best;In terms of Wu Mei printing with multi-font,the comprehensive comparative experimental results show that the SRN text recognition method based on Transformer has the best overall effect.4.The multi-font printed Tibetan text recognition system is developed.The method model obtained from this paper is deployed on the server,and an online multi-font printed Tibetan recognition system in the form of API unified interface and front and back-end interaction is realized.