Deep Learning Based Methods Research On Scene Text Detection And Recognition

Scene text detection and recognition are two important tasks in the field of com-puter vision.Scene text detection aims to locate text instances in the natural scene im-age.Scene text recognition aims to convert a natural scene image that contains only one single text instance into a string that the computer can understand.Compared with tra-ditional optical character recognition,scene text detection and recognition face some challenges such as complex image background,changeable text styles,and imaging quality issues.To tackle these problems,we establish an industrial scene text image dataset and propose two novel scene text detection and recognition methodsFirstly,we establish a scene text detection and recognition dataset in the indus-trial field,i.e.,the Equipment Nameplate Dataset.This dataset contains 502 equipment nameplates images collected in the natural scene.The locations of nameplates,the lo-cations of text instances and the transcripts of text instances are carefully annotated This dataset contains 175 different types of nameplates,and includes several types of characters like Chinese characters,English characters,numbers and symbols,as well as several different forms of text like raised text,engraved text,printed text,handwritten text,so it is very challengingSecondly,to solve the problem that perspective transformation affects the accuracy of text detection,we propose a novel scene text detection method based on key point locating.We design a key point locating network to locate the key points of text re-gion,and transform the text image according to the position of key points.This method not only solves the problem that perspective transformation affects the accuracy of text detection,but also suppresses the interference of complex image backgrounds on text detectionFinally,to solve the accuracy decrease problem of non-horizontal text instances recognition,we propose a shape robust scene text recognition method.We introduce a local direction refinement module to obtain more accurate text control points,and recti-fying text instance images with thin-plate-spline transformation.This method improves the robustness of the scene text recognition model to text images of different shapesWe have conducted a large number of experiments taking nameplate text detection and recognition as examples to verify the effectiveness of the scene text detection and recognition method proposed in this paper.The experimental results show that the key point locating network can effectively solve the problem of perspective transformation affecting the accuracy of scene text detection,and the local direction refinement module can significantly improve the accuracy of non-horizontal text recognition.