| Image based text detection and recognition technology is one of the important tasks in the field of computer vision,optical characters and natural scene text are two important types of recognition targets.Optical character recognition refers to the extraction of text information from scanned document images,and the technology for scanned document text recognition is now relatively mature.The other category is natural scene text recognition,which extracts text from natural scene images.Due to the complex background,poor imaging quality and diverse text styles in natural scenes,it is much more difficult to recognise text in natural scenes than the former,and the current mainstream text detection and recognition methods are not sufficient to meet the needs of practical industrial applications.This paper focuses on the application of natural scene text recognition in the industrial field,taking engine nameplate recognition as an example,the work in this paper mainly contains the following aspects.1.The current research status of text detection and recognition technology in natural scenes is introduced,including target detection,text detection,and research progress in the field of text recognition.2.From the perspective of practical applications,an engine nameplate recognition dataset is constructed,which includes about 1400 images of diesel engine nameplates taken by portable devices,and the text areas in the images are manually annotated,as well as the corresponding text contents.Due to the harsh working environment of diesel engines,most of the nameplate surfaces are rusty and the text contains various characters such as numbers,letters,symbols and Chinese characters,as well as engraved,printed and raised forms of printing,etc.The nameplate images taken under natural scenes also cannot guarantee good clarity,fixed shooting angles and horizontal orientation of the nameplate images.In addition,with the advent of the information age,in engine inspection and maintenance,the information on the nameplate needs to be entered into the information management system.Although nameplate text recognition is a simple and highly repetitive task,it is currently not highly automated and relies heavily on manual recognition,which is inefficient and has limited accuracy in complex scenarios,resulting in high labour costs for enterprises.So this dataset has a strong challenge and practical value.3.In this paper,we design a nameplate text detection and recognition method,which mainly includes two parts:detection and recognition.In the text detection stage,we design an image orientation classification network,from which we obtain the general orientation information of the nameplate image,and then combine the table border line in the nameplate image to adjust the nameplate to the horizontal direction,which solves the adverse effect of image orientation on the text recognition part;secondly,we introduce and improve the DB model to locate the text position in the image.In the text recognition stage,we introduced an adaptive image space transformation module to correct the text image using the thin plate spline trans-formation,solving the problem of text recognition accuracy being affected by image distortion in the text area caused by the perspective transformation;we designed a text recognition model that can make full use of the visual features of the text area and the semantic features of the text itself,improving the problem of individ-ual characters being blurred due to rust on the surface of the nameplate or blurring caused by improper photography.Finally,we introduce the Reed-Solomon error cor-rection algorithm and demonstrate that adding a small amount of error correction information to the text instances during the nameplate printing stage can further improve the accuracy of text recognition.We have carried out extensive experiments on the engine nameplate recognition dataset,and the experimental results show that the method proposed in this paper can meet the needs of industrial scenarios. |
PDF Full Text Request