The Research And Implementation Of Text Recognition System Based On Deep Learning

With the arrival of the information age,picture information is flooded with all aspects of life.The traditional image recognition is based on the surface features of the image.The general methods of extraction have three steps:image segmentation,image feature extraction and classifier recognition.Because text information has no fixed shape and reasonable target boundaries,it is relatively difficult for traditional image recognition methods to identify text information in natural scenes.Deep learning technology is a fast developing technology.Deep learning technology does not rely on the surface features of an artificial image,but is driven by data,learning from the data from the data to the features of the image.The benefit of the data based learning method is that it is not required to set the characteristics by manual intervention.The features of network are more abstract and reasonable.First,a text detection model based on CTPN is implemented in this paper.This method learns the sequence features based on bidirectional LSTM structure by learning the text fraction of small windows,and finally links the text with higher score to the text area.In a natural scene,even if the text does not have a distinct boundary area,it can also detect the position of the text well.The algorithm is tested on 300 Street images of Street View Text dataset provided by Google.The experimental results show that the recognition accuracy of this algorithm can reach 80.2%.It can complete the text detection function in the natural scene very well.Second,a text recognition model based on CRNN is implemented is this paper,which takes into account the advantages of the convolution network to feature extraction and the sequential input of the loop network.Finally,the text recognition system based on deep learning is implemented,which consists of four modules:text detection module,dark channel fog removal module,horizontal normalization module,and text recognition module.The text detection module and the text recognition module depend on the implementation of the corresponding model respectively.The fog module in the dark channel can remove the fog attribute in the image,and the level normalization module can ensure that the text information is in the horizontal position.The accuracy of recognition is improved and the influence of noise is reduced.The integrated multi module is constructed as an image text recognition system.