Scene Text Localization And Recognition Algorithm Research Based On Convolutional Neural Network

The text is the most important information and the knowledge carrier.Natural scene images contain a lot of text,such as road signs,posters,billboards and product packaging,which can convey a lot of semantic information to people.The target of text recognition in natural scene images is to accurately locate and identify the text information,which can propose the deeper understanding of the scene.As for the complex background and lighting conditions,as well as multi-scale,multi-lingual and multi-directional,the efficient text recognition in natural scene images is a challenging task.CNN is a learning model for feature extraction under the supervision mechanism.As for it can use images as input directly,CNN has natural advantages in image processing.Based on the study of CNN and related techniques of text location and recognition in natural scenes,a natural scene text detection and recognition method based on CNN is proposed in this paper.The main contributions of this thesis are as follows.1.A text region location method based on edge-enhanced MSER and CNN is proposed.This method as for the blurred edges of adjacent characters,the edge-enhanced MSER regions are detected as character candidate regions.The non-character areas are deleted using five features.The CNN is used to delete non-character areas that geometry features cannot delete.And the resulting character regions are combined into a region capable of expressing semantic information words through positional relationships.2.Realize the character recognition method based on convolution neural network.The method first implements a character classifier,This paper comparing the influence of CNN with different depths,different convolution kernel sizes,and different learning rates on character recognition.Get the image of each character from the results of the text location,the normalized character images are sent into the character classifier for recognition.Then combine the recognized character according to the position of the word to get the result of word recognition.The edit distance algorithm is used to correct the recognition result.Finally,this paper uses the ICDAR 2013 database to establish a data set and the proposed method is evaluated on the ICDAR 2013 Dataset.And the experiment results show that the proposed method can achieve good performance.