Natural Scene Text Recognition Based On Deep Neural Network

Text,as the tool of recording thoughts and carrying language,plays a pivotal role in the development of human society.The text in the natural scene is full of information.Using modern technology to obtain the information can bring great convenience to human work and life.The text in the natural scene is complex and changeable.Compared with the printed text which has a typesetting standard and high resolution,it has the characteristics of various fonts,complex background,random distribution and many interference factors.The recognition rate is low when using traditional optical character recognition technology(OCR technology),cannot met the actual use requirement.Based on the related literature research in the field of natural scene text recognition and deep learning in China and abroad,this paper proposed a natural scene text recognition method based on deep learning to recognize the natural scene text,achieving certain recognition accuracy and effectiveness.The main innovations and research results of this paper are as follows:(1)An end-to-end deep neural network framework combining convolutional neural network(CNN)and recurrent neural network(RNN)is proposed.Text recognition in natural scenes is divided into two parts: feature extraction and feature recognition.The model adopts an encoder-decoder structure,a convolutional neural network is used as an encoder of the model to extract feature vector information of the input text image.The recurrent neural network is used as a decoder to identify the front and back inputs.The final model obtained after training was tested on four popular and representative data sets IC03,IC13,IIIT5 K,SVT.The results show that the designed model has superior algorithm and high recognition rate,and is suitable for text recognition in natural scenes.(2)A soft attention mechanism is added to the deep neural network for natural scene text recognition,which is located between the convolutional neural network and the recurrent neural network.The experimental results show that the attentio n mechanism can be used to further extract the feature vectors that are favorable for the output,improve the output accuracy.The natural scene recognition model combined with convolutional neural network,recurrent neural network and attention mechanism can train according to the word level annotation.It does not depend on the fixed dictionary,and can be recognized without pre-processing the input image,achieving end-to-end depth neural network training,which means it has good universality.