Scene Text Segmentation Based On MSER And Adaboost

Natural scene text images contain a large amount of text information.The automatic extraction of these text information is of great significance for scene understanding,environment cognition and content-based image retrieval technology.The text extraction system in natural scene images is divided into three parts:text detection,text segmentation and text recognition.In order to extract the text information in natural scene image,we first need to detect and locate the text from the complex background.Secondly,the detected text should be separated from the background to facilitate identification.However,scene text is easy to be affected by uneven illumination,complex background and other factors,which results in poor quality of text detection.If the detected text area goes straight to the recognition process,the recognition ratio is reduced considerably.An effective algorithm for segmentation of scene text based on maximally stable extremal regions?MSER?and AdaBoost is proposed to overcome the interference of uneven illumination and clutter background to scene text segmentation.At first,according to the input image,the preprocessing step is divided into two polarities?bright text on dark background and dark text on bright background?,and then the R,G,B three color channel images are obtained on the base of two polarities,which can be divided into six images.Secondly,the MSER algorithm is selected to extract text candidates according to the text of the natural scene image after preprocessing.Furthermore,we select some heuristic rules and stroke width transform?SWT?rules to carry on text selection.Then,in the process of character classification,we use an AdaBoost-trained classifier that adopts the texture feature of uniform mean local binary patterns?MLBP?to divide each character candidate into either character or non-character,and the regions that are not text are eliminated.Finally,the final text segmentation output is obtained by combining the results from the R,G,B color channels in two polarities.The proposed method is evaluated on the ICDAR₂013 datasets and experiments show that it performs well and can achieve good segmentation results.