Research On The Detection And Recognition Method Of Wujin Tibetan Script In Natural Scenes

Text detection and recognition have very important research and application value.Traditional text detection and recognition technology has good performance on printed text,but the accuracy of text detection and recognition in natural scenes is still very low.How to narrow the performance gap between text detection and recognition in natural scenes and traditional text detection and recognition is an urgent problem to be solved in text detection and recognition in natural scenes.At present,the detection and recognition of Wujin Tibetan script in natural scenes is in the initial stage of exploration,and Wujin Tibetan text information in natural scenes,as highly condensed high-level semantic information,not only has great research and practical value,but also can be used for assistance.Research in the field of Tibetan contextual text understanding.There is a significant difference in the background between the Wujin Tibetan text image of the natural scene and the scanned Wujin Tibetan text image.The background of the scanned Wujin Tibetan text image is simple,usually a single white color with little interference.The background of Wujin Tibetan text images in natural scenes is complex and varied.At present,there are few related researches on the detection and recognition of Wujin Tibetan text in natural scenes.Effective detection and recognition of Wujin Tibetan text in natural scenes has great practicality value.This article discusses the research background and significance of the detection and recognition of the Wujin Tibetan script in natural scenes.It describes and analyzes the current research status of Chinese,English and Tibetan text detection and recognition technology in natural scenes,and analyzes the structure of Tibetan characters.And the rules for the composition of Tibetan characters are introduced.Based on this,the research is focused on the detection and recognition algorithms of Wujin Tibetan script in natural scenes.The main work of this paper is as follows:1.In the early stage of collecting the data set,first analyze the natural scene text data set and image characteristics,and then collect and label the Wujin style Tibetan text detection data set and the recognition data set in the natural scene;in the later stage of collecting the data set,first use the network The crawler technology collects a large number of Tibetan corpus and preprocesses the Tibetan corpus,and then uses Open CV-Python to synthesize the Wujin Tibetan text block,which solves the problem of insufficient image data of the Wujin Tibetan image in natural scenes.2.Study the detection method of Wujin Tibetan script in natural scenes,adopt EAST and DBNet algorithms,and replace the original feature extraction network,and compare the detection performance of different feature extraction networks such as Res Net18,Res Net34,and Res Net50 on the test set.3.To study the method of recognizing Wujin Tibetan script in natural scenes.Firstly,based on the study of the structure of Tibetan characters and the rules of Tibetan character formation,design the whole Tibetan character generation algorithm,and generate the whole Tibetan character set as the recognition of Wujin Tibetan script.Dictionary;then study the preprocessing method of the Wujin Tibetan image block of the natural scene to be recognized;finally study the sequence-based text recognition algorithm CRNN,and use the improved Mobile Net V3 Large as the feature extraction network,and compare it under different data set sizes.The recognition accuracy of the algorithm on the test set.Based on the existing research foundation,this paper has achieved the following results:1.Collected and annotated the natural scene Wujin type Tibetan text detection and recognition data set,which contains 1796 natural scene Wujin type Tibetan images,4321 natural scene Wujin type Tibetan text image blocks;collected 6,284,247 Tibetan short sentences,Synthesizing200,000 image blocks of Wujin Tibetan text with annotations;2.The detection of Wujin Tibetan script in natural scenes is realized.The results show that the DBNET algorithm using Resnet18 as the feature extraction network has better detection performance in the actual test.The accuracy,recall rate and F1 value of this method in the test set reach 0.87,0.55 and 0.67,respectively;3.The recognition of Wujin Tibetan script in natural scenes is realized.The experimental results show that the CRNN algorithm using the improved Mobile Net V3 Large as the feature extraction network has a recognition accuracy of 0.7301 on the test set.And on this basis,it analyzes the special examples of errors in the recognition of Wujin Tibetan script in314 real natural scenes,which will provide a certain reference for related researchers in the future.