Research On The Application Of Convolution Recurrent Neural Network In Scene Text Recognition

In this world,information can be said to be everywhere,and image,as one of the most important carriers of information,is able to enable people to obtain a lot of valuable information.In recent years,with the continuous development of deep learning,character recognition technology has become more and more mature,and related applications have also emerged,such as ID card recognition and license plate recognition.These traditional OCR(Optical Character Recognition)technologies have one common feature,that is,the Text format of the object to be recognized is relatively uniform.However,the current society has an increasingly high demand for Character Recognition in complex scenes,which derives STR(Scene Text Recognition)technology.The purpose of this topic is to use STR technology to detect and identify the shop signs of the merchants along the street,so as to make the information collection of the merchants by the Pudong urban management team more accurate and efficient.Scene text recognition algorithms can be generally divided into two main parts : text area detection and text recognition.Based on the application scenario of merchant sign recognition,this paper proposes a store sign recognition model,which is mainly composed of Pixellink and CRNN neural networks.Pixellink network is used for text area detection.Based on the distribution characteristics of store signs in the real scene,a subject instance campaign algorithm is designed in this part to exclude the secondary text in the picture.CRNN(Convolutional Recurrent Neural Network)is used for word recognition,and in this Network,the efficiency of model training is improved by increasing the number of batchrecycling Neural Network.Finally,there is a text similarity comparison module to improve the accuracy of text recognition.This module uses Levenshtein distance algorithm and text2 vec based method to correct English text and Chinese text respectively.In the test set,a total of 60 photos were taken with Chinese name and a mixture of Chinese and English name respectively.The recognition accuracy of the model in this paper reached 56.7% and 58.3% on the two test sets,and the ablation experiment showed that each functional module played a positive role in improving the model recognition ability.It is concluded that the model in this paper can better accomplish the identification task of the shop sign of the merchant along the street after further improvement.Later,this study will be applied to the street-side merchant inspection module of " Zhi Fa Tong " APP of Pudong Urban Management.This functional module aims to check the environmental problems of merchants,such as whether dry and wet garbage bins are placed in front of the door,and whether there are noise pollution,lampblack pollution,sewage discharge and other problems.