Research On The Algorithms Of Text Character Processing In Natural Scene

The text in the natural scene is widely present in the road signs,billboards,license plates,all kinds of instruments.Character detection and recognition has become a hot research topic in the field of computer vision and document analysis.However,the detection and recognition of text in natural scenes is a challenging problem.The difficulties are mainly due to the diversity of the text,the complexity of the background and the interference factors in the imaging process.If the original natural scene images are sent into the text recognizer directly,there will be a lot of problems,such as identification errors,missing text and no identification results.Therefore,the accurate text detection and localization algorithm will greatly improve the accuracy of text recognition and semantic understanding in natural scenes.In this paper,the method of text character processing in natural scene mainly aims at text detection in natural scene.The purpose of this paper is to locate the text in the natural scene accurately and guarantee the reliability of the subsequent character recognition.The main contents of this paper are as follows:1.In this thesis,we study the method of character detection based on MSER.Aiming at the problem that the algorithm can not detect the text in the lower contrast image,this thesis proposes a method of multi-color space MSER character detection based on Retinex.This method enhances image contrast and brightness by the Retinex enhancement algorithm and extract the MSER of grayscale image and the luminance channel image in HSI color space.And then,the maximally stable extremal regions in the two images are combined for the text candidate regions.After testing,the improved algorithm can improve the performance of detection effectively when the image is affected by the light intensity.As a result,the application scope of the algorithm of character detection based on MSER is enlarged.2.This thesis studies the common method of character processing in natural scenes.Aiming at the problem that the single detection algorithm has a poor effect on the image with complex background,this paper proposes a character detection method based on the fusion of MSER and SWT.In this method,the connected region is obtained by calculating MSER and SWT firstly,and then the candidate regions are obtained based on the distance relation of the connected regions extracted by the two.The test results show that the method can effectively improve the performance of character detection in complex background.3.This thesis analyzes the results of character detection in the image with complex background.Aiming at the problem that false alarm rate in the results is too high,the thesis proposes the method of text regions verification by the classifier.In this method,the random forest decision tree is used to train the samples,and then the trained classifier is used to carry out the verification of the detected text regions.After testing,the improved method can reduce the false alarm rate effectively.4.In this thesis,we analyze the features that are used for training the classifier.Aiming at the problem that single feature can not achieve good classification results,this thesis proposes a method that fuses the multi features to form character features.The method extracts the HOG features and the LBP features of the text region respectively,and then obtains the feature vectors by fusion the two features with a serial way.After testing,the improved method can improve the performance of classifier effectively.