Design And Implementation Of Scene Text Recognition System

In recent years,with the advent of deep learning and convolutional neural networks,the development of image processing technology has made a qualitative leap.More and more traditional methods have been replaced by methods based on deep learning,such as text recognition.While text recognition for natural scenes often focuses on specific areas,such as license plates,street scenes,etc.,there are still many application problems for text recognition in natural scenes without good solutions.In real text scenes,the performance of Chinese character detection and recognition is affected by background light,angle,and data set differences,and the results are unsatisfactory.This subject analyzes the current research status of domestic and foreign text recognition technologies and the disadvantages of text recognition systems.Based on this,a set of general frameworks for scene text recognition based on deep learning technology can provide image recognition functions in natural scenes.OCR system lacks identification of card scenes,and each identification system can only identify a single type of document.Based on the original OCR technology,it integrates the identification function of universal documents,including ID cards,bank cards,business licenses,and The recognition of driver’s license has realized a multifunctional scene text recognition system.It also provides an open API for function expansion.At the end of the article,the function is extended with a business license identification.Unicom Tianyan API provides extended enterprise information verification functions to implement and verify open capabilities.Aiming at the design and implementation of scene text recognition system,the main contents of this thesis include the following four aspects:1.A method for synthesizing scene text datasets is proposed:By performing background analysis and feature extraction on document images in real and natural scenes,the image influencing factors such as background texture and lighting angle are simulated and realized,and the Chinese coverage rate is as high as 99.7%Scene text character library,thus increasing the generalization effect of each algorithm in the entire detection and recognition process.2.The scene recognition method of IBN-CRNN is proposed:Improve the current advanced end-to-end trainable variable-length text recognition method CRNN.The introduction of the IBN-Net structure can help improve the accuracy and accuracy of the model without increasing the amount of calculation.Generalization,combined with synthetic data sets,can effectively identify image text in interference background.3.A general framework for scene text detection and recognition based on VGG+CTPN+IBN-CRNN is proposed:Utilizing the characteristics of CTPN suitable for detecting horizontal text,the VGG 16 algorithm is introduced to classify images at oblique angles to achieve image correction;combined with the synthesized data set and The open data set enriches the data training samples for detection tasks and recognition tasks.Training on the proposed IBN-CRNN scene recognition model using the above-mentioned expanded sample data enhances the generalization of the IBN-CRNN algorithm and the robustness of the overall framework to the Chinese scene.4.Designed and implemented a scene text recognition system:The system encapsulates capabilities based on scene text detection and recognition universal frames to provide a universal text recognition service with high accuracy in natural scenes.The system also integrates common card image recognition functions,including identity Identification,bank cards,business licenses,and driver’s licenses.The system provides a friendly programming interface to easily expand new functions for specific application scenarios.It has also realized the recognition of business licenses and the Unicom Tian Eye Inspection API provides extended enterprise information verification functions.The system provides an interactive and responsive Web interface for access by different terminals.The interface supports image uploads and web image address.It also supports one-click copying and editing of identifying content.The system can be deployed to ensure data security in an intranet environment.