Detection And Recognition Of Scene Text In Complex Background Based On AdaBoost

The text information in the natural scene image is of great value and can be applied to many aspects such as image retrieval,driverlessness,and visual aids.It is a hot research topic in the field of computer vision.The extraction of natural scene image text is mainly divided into three parts: character positioning,character merging,and character recognition.Character positioning is the key step of extracting text,which has a significant impact on the recognition accuracy of the following characters.Although the traditional OCR technology has been very mature,the recognition accuracy of document image texts has reached nearly 100%,if it is applied directly to the text recognition of natural scene images,the effect will be greatly reduced.Because the document images generally have high resolution and clear text.The contrast between the background and the characters is high,and the tilt deformation does not occur substantially.However,natural scene images often contain a large number of complex background textures,and are affected by illumination,fonts,and shooting angles,making the positioning and recognition of the text face many difficulties.This dissertation studies the techniques of text detection and recognition in natural scenes,and proposes an algorithm for detection and recognition of natural scene image text candidate regions based on the maximum stable region MSER and the adaptive algorithm AdaBoost.The main work includes the following aspects:(1)According to the characteristics of natural scene images,a pre-processing method that can improve the efficiency of natural scene image text acquisition is designed.Preprocessing includes graying,smoothing,and sharpening.Grayscale can simplify the image,reduce the computational complexity of the algorithm,and improve the processing speed of the algorithm.The Gaussian blur is used smoothly.After smoothing,the noise of the image can be removed.For sharpening,the Laplacian sharpening operator is used,and sharpening can be enhanc ed.The edge of the image highlights the details of the image.(2)A method for extracting candidate regions of natural scene image based on the maximum stable extreme region is designed.The method finds the extreme regions in the image by increasing the binarization threshold and calculates the rate of change of these extreme regions.Text candidate area.(3)According to the characteristics of the text candidates,we propose three methods for screening text candidates: heuristic rule filtering,stroke wi dth filtering,and Ada Boost-based classification model filtering.The heuristic rule filtering is based on the candidate area area and the duty cycle feature to set the filtering rule.The stroke width filtering is to extract the stroke width of the text in the candidate area and delete the area that does not meet the width of the stroke;the classification model based on Ada Boost needs to be firstly needed.The gradient and texture features of the candidate region are extracted by using the gradient histogram and the local binary model,and the classification model is obtained by inputting a weak classifier composed of a decision tree.The candidate region is divided into a text region and a non-text region,and ICDAR-2003 is adopted.The image set was simulated and the simulation results showed that the text in the natural scene image can be located accurately.(4)A character merging method and a character recognition system based on Tessact are designed.The expansion method of mathematical morphology con nects the similar characters in the image together,and the connected characters are merged through the connected domain analysis,and Tesseract is called to perform character recognition.Finally,a simulation experiment was conducted.The experimental re sults show that the system has a high recognition accuracy and has a certain theoretical significance and practical value.