Chinese Text Recognition On Mobile Terminal Based On Deep Learning

In recent years,with the popularity of shooting equipment,image acquisition has become simple.In a simple image can contains tons of information.The way of data transmission and storage has gradually changed from the past text format to the new media format such as image and video.The information in the image is displayed by the way of pixels,which makes the user understand more intuitively.For the computer,it is necessary to understand the information contained in the image Content is of great significance.The text content in the image has a high degree of identification,and has a strong correlation with the content shown in the image.The high-level semantic information contained in the text information plays an important role in the understanding of the image.At the same time of the development of artificial intelligence algorithm,it also needs enough data to support.Especially,deep learning has gradually become the mainstream method.Supervised learning algorithm relies on data feature processing to provide machines for training.It is expensive and inefficient to label data sets only by manual methods,and a large amount of image content is difficult to be utilized.Labeling text data has become an important problem to be solved before algorithm research.Due to the diversity of scenes in which text appears,it is necessary to implement various methods of text labeling.There are a lot of text content in daily life,such as menu,flyer,billboard,etc.by converting the image into text information that can be understood by the machine,it can be further operated,such as translation,information statistics,etc.,which provides convenience to the user.1.Text localization in natural scenes,combined with the relevant knowledge of target detection,the end-to-end target detection algorithm and R-CNN series algorithm are studied.This paper designs a text localization algorithm using YOLO-v3 structure,and makes a comparative experiment with CTPN algorithm.The results show that its localization accuracy is lower than the latter,but it has an absolute advantage in computation time.2.For text recognition tasks in natural scenes,the CRNN text recognition algorithm combined with RNN structure is analyzed.The algorithm uses CNN to extract spatial features and RNN to extract contextual features,which can better complete text recognition tasks.Through model training and experiments,it is proved that CRNN network can complete text recognition tasks in natural scenes.3.Combining text location and text recognition,this paper designs and implements a text annotation tool,which can adjust the annotation content through manual correction,greatly improving the efficiency of dataset annotation.