Natural Scene Text Localization And Recognition Based On Mobile Terminal

With the widespread popularity of mobile terminals,especially smartphones with camera functions,people can easily capture a large number of natural scene images.At the same time,along with the rapid development of deep learning and computer vision,natural scene text localization and recognition and how to transform existing neural network models into mobile terminals have become the research hotspots in recent years.The research in this direction has important theoretical significance and broad application prospects.This paper takes the natural scene as the research background,focuses on how to efficiently identify text information from the natural scene and compress the deep neural network model,and designs and implements the natural scene text recognition system based on the mobile terminal.The main contents of the paper are as follows:(1)In the natural scene text localization,this paper starting from the target detection problem,after researching and summarizing the excellent text localization algorithm existing at this stage,this paper designs a text localization algorithm based on YOLO v3 structure,and compared with existing CTPN algorithm has carried on the experiment,the experimental results show that YOLO algorithm on reasoning time-consuming than CTPN sharply reduced,but in the former than the latter scenario text localization accuracy.(2)Aiming at the text recognition problem of the natural scene,this paper analyzes a CRNN text recognition algorithm composed of CNN and RNN.Through the pre-training of the model and a series of evaluation experiments,the results show that CRNN is capable of text recognition task of a general natural scene in the case of a small model scale.(3)For the deep neural network model compression,this paper first analyzes the difficulty of porting the existing network model to the mobile terminal,and then discusses the current CNN and RNN compression methods,and combines the CRNN algorithm to carry out the model compression experiment.The results show that at the expense of a certain accuracy,the compressed model can meet the needs of real-time reasoning in the mobile terminal.(4)Combined with the above three parts of research,this paper designs and implements a natural scene text recognition system based on the mobile terminal.