Study On Text Detection Method Of Complex Background Image By Integrating MSCRs Into MSERs

The text in images contains valuable information,such as signposts,license plates,advertisements,trademarks and posters,which can be exploited in blind assistance,automatic navigation,unmanned driving and many other multiple application areas.The existing image text extraction technology can detect and recognize the text in document image with simple background,whose size and color is consistent.However,location of the text area should be specified for text recognition under non-uniform lighting and complex background conditions,otherwise the recognition result is poor.More and more applications need to extract text information from the complex background images along with the rapid development of mobile Internet technology,such as mobile devices,wearable equipment,unmanned vehicles,etc.The study of fast text detection in complex background image is one of the core problems in these applications.Method based on the sliding windows and method based on connected component(CCs)are two most popular image text detection methods at the present time.In particular,Maximally Stable Extremal Regions(MSERs)based method is a typical representative of the CCs based method.The MSERs algorithm is scale and rotation invariance,which can be adapted to texts with different size,angle,and color.However,the MSERs algorithm is sensitive to the change of text background,which make it performs poorly when used for text detection under complex background condition.In this paper,based on the analysis of the reason why the MSERs algorithm is sensitive to the change of the text background,a novel text detection method for complex background image is proposed by integrating MSCRs into MSERs.First of all,we take the region extraction results of MSCRs and MSERs as candidate characters,which make sure that our method can extract more real characters than MSERs algorithm alone.Then,to remove the non-character regions,we train a random forests character classifier according to the texture features extracted from the character and background samples.At last,the scattered characters are constructed into text regions according to the color and spatial location features.Our method can successfully detect the text with variety of background,scale,angle and color in complex background images according to the experimental results.In the ICDAR 2013 database,out method get 71.9%in recall rate,84.1%in accuracy rate and 77.5%in F value,and the recall rate and F are higher than the compared methods.