Research And Improvement For Algorithm Of Text Detection In Natural Scene Images

Text detection is a necessary prerequisite for many text recognition applications.With the 'increasing use of mobile phones and other facility,images have become a carrier to record man's life,especially natural scene images.Although OCR technology has been very mature,it is oriented mostly for document images when it comes to text detection in images.In document images,text background is single and can be separated with words easily.However,natural scene images have various types of text and complicated background,which makes it a big challenge to detect text.Text detection methods based on maximally stable extreme regions(MSER)are popular in the field.After conducting a study and analyzing the MSER based methods,we introduce new classify features and deep learning method,which improves detection rate effectively.Main results are as follows:Firstly,we.propose a moment and texture based text detection algorithm.Classifiers incorporated with new geometrical and texture features show better accuracy.In the stage of pruning letter candidates,the algorithm uses a character classifier incorporated with Hu moment features to delete most non-character candidates;a text classifier mainly trained with symbiotic moments is used in the stage of deleting text candidates,improving the recall rate effectively.To further improve the detection accuracy,an algorithm based on deep learning is improved on the basis of the formal algorithm.As deep learning method has high classification accuracy,the deep belief network is applied in the stage of deleting text candidates.To improve the quality of training images,median filter algorithm is used to remove noise.After comparing with other methods,both based on MSER and not,our method performs well in precision and recall rate,resulting in better f-measure.Above all,the method for text detection in natural scene images here is an effective one.