Natural Scene Text Detection And Recognition Based On Deep Learning

With the rapid development of multimedia technology,the technology of intelligently obtaining information from big data has broad application prospects.The information contained in the text is directly and efficiently.Digitizing these textual information is of great significance for improving the multimedia retrieval capability,industrial automation level,and scene understanding ability.Nowadays,there are lots of mature applications of document text recognition technology.Detection and recognition of texts in image,especially in natural scene image,is an important but challenging task.It is still at the stage of research and exploration.The main research contents of this thesis are as follows:1.This thesis studies a fast natural scene horizontal text detection algorithm.Firstly,a feature extraction network for horizontal texts of natural scenes is designed.In this thesis,a concatenation network structure with multiple convolution channels is used,and a convolution kernel conforming to the text line shape features is used,and the middle layer features are merged and enhanced.Then the horizontal text feature extraction network designed in this thesis is applied to a regression-based detection model.The number of detection branches is reduced and the network model parameter scale is reduced.An end-to-end fast natural scene horizontal text detection model is obtained.2.This thesis studies an accurate natural scene multidirectional text line detection method.Inspired by a detection method which is based on region proposals,this thesis combines the context information of text lines to improve the grouping and connection of detected text proposal boxes.First,according to the relative position between the proposals,it is determined whether to divide them into the same text row construction group;then,the tilt angle is fitted according to the center position of the same group of proposals;finally,the final text line detection result is selected by using a rotating rectangular box.The detection algorithm makes use of the advantages of the region proposal based detection algorithm in terms of accuracy and lays a foundation for the subsequent text recognition algorithm.3.This thesis studies a natural scene text recognition algorithm based on residual network.This thesis uses text recognition method based on residual network to explore the role of the recurrent neural network in text sequence recognition.In order to train and test the text recognition network,this thesis constructs a database for text recognition in natural scenes,especially for the identification of Chinese texts.Finally,the text detection and recognition algorithm of this paper is tested on the real natural scene text image.The experimental results show the effectiveness of the introduced algorithm.