Research On Text Detection Technology In Natural Scene Pictures

With the development of smart phones and the Internet,it is simper and faster to get and share scene images.The text in the scene images contains rich and accurate information,and the detection of text information is important for understanding the image.In the meantime,text may be the focus point of users.Text detection in scene images refers to determine whether the picture contains text,if it contains text,then separate text area with background.Detected text area is useful for subsequent translation,text recognition or speech processing.Scene text detection remains many challenges,such as complex background,the great diversity of text itself,the poor quality of images etc.Traditional algorithms based on connected components often comprise with character candidate detection,text line grouping,and text line filtering steps.These approaches remain these following problems:it's hard to filter out character-like noise in character candidate detection step,and some true text area may not be detected.It's difficult to detect text lines with one or two characters for the lacking of text line constraints;without considering the detecting of the arbitrary orientations text.Based on these reasons,this paper presents a method based on cascaded filtering and grouping modules,and modify it to detect arbitrary orientations text.The detailed work of this paper is as follows:1)We combine feature based ER method and aspect ratio constraint based MSER method to extract character candidates.In the text grouping stage,we propose a distance and similarity based character pair linking module.We also joined the single and double character detection module and non-text filtering module based on the similarity of characters and the statistical feature of text lines to our pipeline.This method can improve the precision and recall of character candidate detection and compensate the loss of character detection,and have good perfomance with filtering out background area.Experiments on the ICDAR2011 and ICDAR2013 datasets demonstrate that our method can improve the detection recall and precision.2)We extend our work to detect arbitrary orientations text.We analyse the similarities and differences of horizontal text detection and arbitrary orientations text detection,and change the exsiting horizontal text line grouping method to arbitrary orientations text detection task.We using center point of the character candidate bounding box as feature to detect arbitrary orientations text.And we rotate the character candidate bounding boxes and text line bounding boxes to detect arbitrary orientations text.Experiments on USTB-SVT dataset show that our method can effectly detect arbitrary orientations text.