Detecting And Recognizing Text In Natural Scene

It's meaningful for human beings to detect and recognize text existing in pictures and videos,which will play an important role in the area of scene recognition,infor-mation retrieval and so on.However,the text in natural scene is heavily effected by many factors,such as cluttered background,lighting condition,prospective distortion and blur.Therefore,it's a challenging task to detect and recognize text in natural scene.To over come these difficult,this paper will propose a solution of end to end natural scene text recognition.The solution of end to end natural scene text recognition consists of three parts:text detection based on similar maximally stable extremal regions(MSER)grouping,character recognition using markov random field(MRF),and word recognition based on bigram.The method of text detection extracts MSER as candidate character region-s,then groups candidate character regions to form text line based on the similarity of them.The method of character recognition builds an undirected graph representation for each character category and regards the procedure of character recognition as the procedure of projecting undirected graphs to test image.The projection procedure is restricted by energy function of MRF.The method tries to describe the shape of char-acter from the prospect of local feature and global structure.It's showed respectively in the unary item and pairwise item of energy function.The method of word recognition is used to deal with wrong recognition result of single character in order to improve word recognition rate.This method utilizes the recognition result of single character and context of characters in the same time.We have done comparative experiments of text detection,character recognition and end to end text recognition on several widely used datasets.The experiment results show that the approach proposed in this paper is feasible for detecting and recognizing text in natural scene and competitive.