Design And Implementaion Of Deep Earning Based Mongolian OCR System

With the development of intelligent technology represented by artificial intelligence,more and more organizations and individuals are aware of the opportunities and challenges brought by artificial intelligence.Among them,OCR technology,uses machines to automatically recognize text on images and uses them for subsequent tasks of storing,transferring,processing,and application,plays an important role in technologies such as autonomous driving.With the development of deep learning technology,the accuracy and stability of the OCR system have been greatly improved by using the deep learning model.At present,the research and development of OCR technology for Chinese,English and other popular languages have been very perfect,and it can obtain the recognition accuracy close to human beings,but OCR technology for minority languages still has a lot of room for development.This paper mainly studies the OCR technology for Mongolian.This paper systematically investigates the research and application in the field of Mongolian OCR,and finds that the current Mongolian OCR system can only handle text recognition tasks in relatively simple scenarios,such as text recognition tasks without background interference.In real life,text recognition scenarios are diverse.Pictures may contain background patterns,noise,lighting,etc.,and text may have multiple fonts,colors,and sizes.Aiming at this problem,this paper first analyzes the requirements of the Mongolian OCR system.Then design the front-end and back-end functional modules and interface definitions of the system.Then research and design the two core algorithm models of OCR services,the text detection model and the text recognition model,in which the text detection model optimizes the feature fusion module on the basis of the text detection DB model;the text recognition model adopts sequence to sequence recognition model CRNN as architecture,and the backbone network has made corresponding structural adjustments based on Mobilenetv2 for Mongolian features.Then build a large number of synthetic data sets and a small number of human-labeled data sets,and train and test models on them.Finally,a Mongolian OCR system based on deep learning was implemented.The main features of this system are:the algorithm is based on deep learning technology,and completely driven by data;the system can recognize OCR tasks in relatively complex scenes,such as multiple fonts,lighting,background,blurry,etc.In terms of system design,each module of the OCR system adopts micro-service design.According to different calculation volumes or business needs,each module can be independently scaled,and can be independently deployed on multiple different servers and operating systems.In summary,this system is an accurate and efficient Mongolian OCR system that is cross-platform,modular,easy to maintain and manage,can adapt to different scales of business volume,and can handle various text recognition scenarios.