A Method For Detecting And Recognizing Non-linear Text

Human civilization has developed to today,the text has always been an important carrier of human communication and knowledge acquisition,text detection and text recognition always occupy a very important position in the field of computer vision because of the development of computer technology.Now most text detection methods are for linear or near-linear text detection,and these methods have good performance for linear text,but a lot of times the text area that needs to be tested is not linear or nearly linear,and in many case,the linear text always becomes curved,trapezoid or irregular due to the design of the character or the angle of the shot,at this time the method used in the past is not succeed for text detection,or the detection result contains a large number of non-text areas,not only will the text field box be imprecise,but it will seriously affect the accuracy of the recognition results in the following text recognition work.This paper presents a method for text detection and text recognition based on text shape in view of the above problems,adopts the latest and popular method of deep learing.First,the dataset MSRA-TD500 is used to train the convolutional neural network and predict the approximate position and direction of text of image,then the region feature extraction is used to obtain the exact location and size of a single character and frame the individual characters using a rectangular box,finally,a text box with irregular shape on the edge of the whole text area is processed precisely for the rectangular frame of single character,and accurate text detection and recognition are realized.In this paper,some images in the NEOCR dataset are used for testing,and compared with the two methods which are better now,and the results prove that the text detection method in this paper is robust to detecting a variety of shapes text area,especially in the image of non-linear text or non-rectangular text area.In addition,the experiment proves that this paper can also get good test results for straight-line text.In order to facilitate text detection and character recognition,this paper uses the JAVA language to implement a system for detection and recognition of non-linear text.