Research On Text Detection Algorithm Of Natural Scene Based On Bayesian Model

In natural scene images,text information is an important information to understand the content of the image.Text detection is a necessary step in image content analysis,and it is also the basis of text detection and recognition system.Because of the complex background and the various forms of text in natural scenes,it is still a great challenge to detect the text information in natural scene images.Therefore,it is of great practical significance and application value to study a robust text detection algorithm.In recent years,the Maximally Stable Extremal Regions(MSER)extraction operator has been widely used in the field of target detection based on the performance of affine invariance.The traditional MSER method,based on the gray level image to extract the extreme area,has the limitation of extraction of extreme values due to the phenomenon of over fusion,which is caused by the gray level similarity of the pixels in the gray image.In order to improve the accuracy of text detection,an improved MSER method was proposed in this paper,which was used to acquire the text candidate region,then the method Naive Bayesian phase was used for further screening of candidate text region.The main work of this paper is as follows:1)In this paper,the principle of MSER feature detection and the process of extreme value region extraction are deeply studied,and the traditional MSER method is improved.Because there are some characters lost or conglutination in the results of the traditional MSER method,this paper adopts an edge-preserving MSER improved method.Experiments show that the candidate regions are better in terms of quantity and quality.At the same time,in order to make better use of color information of image and provide more rich text candidate regions,an edge-preserving MSER improved method on HSI color space,based on three channels of H,S and I,was proposed in this paper.Test results show that this method can get more rich text candidate regions.2)In order to realize further screening of candidate regions,In this paper,we studied some obvious characteristics of text and non text and Naive Bayesian model.According to the text feature,three features of stroke width(SW),color perception differences(CPD)and edge gradient feature(eHOG)were extracted.On the training set of ICDAR2013 data sets,it uses the naive Bias to learn to get the text and non text feature distribution;On the test set of ICDAR2013 data set,the posterior probability of a characteristic is obtained based on Bayesian formula.3)The graph cut algorithm was studied and presented finally.A minimum energy model,which is constructed by the characteristics of the posterior probability,and the characteristics combined with stroke width color difference,and the maximum flow minimumcut theory is used to make a two value mark for the candidate text regions,then region merging and the text regions are comfirmed.Finally,for the final text marking region,the text line is constructed based on the mean shift framework.In this paper,we test the algorithm on the test set of ICDAR2013 text detection game data set,and the experimental results will be submitted on the open platform ICDAR2013 provides.The results show that the proposed algorithm can effectively extract the text region in most natural scene images.