| The text recognition problem has two difficult points.The first is the forms of existence of the text are ever-changing.In particular,Chinese characters are more diverse in strokes,and there are various fonts,etc.The second is that the text may be occluded and the background is more complicated in practical application scenarios.Therefore,two key technologies are involved in doing a good job of text recognition algorithm: Firstly,to extract a well-characterized feature text and an effective classification decision network;Secondly,to design a reasonable data enhancement algorithm and to transform the training samples into images in various situations close to realityl,in order to learn a more robust recognition model.In this paper,a text recognition model is designed based on CRNN model,which successfully realizes high accuracy recognition of various characters.The main work of this paper consists of the following three parts: The first is the design of data enhancement algorithms.This paper designs a suitable data enhancement algorithm to help the network learn better.The second is the design of feature extraction network.This paperbased on the CRNN model and improvethe feature extraction network.The feature extraction network designed by using the Dense-Block structure can not only extract the sample feature information,but also learn the contextual information in the text.The third is the design of the network decision-making layer.In this paper,a convolutional layer is used to replace the LSTM layer which has large parameters and difficult convergence in the original CRNN model.This not only improves the recognition rate to a certain extent,but also reduces the parameters of the network and speeds up the network convergence.In this paper,the Python language and Keras framework programming are used to realize the whole identification system,including data enhancement algorithm,feature extraction network and decision layer of network.The experimental results show that thesystem can recognize a variety of characters well and has high recognition accuracy.The recognition rate of on the Chinese character database is 95.28%,the recognition rate on the English character database is 96.47%,and the recognition rate on the digital database is 98.85%. |
