Research On Text Location In Natural Scene Images

Text location and recognition are the two important modules of text information processing in scene images.As the prerequisite of text recognition,text location has practical application value and theoretical research significance.After the development of last decade,the technology of text location in scene images has achieved certain progress.However,with the complexity of scene images and the interference of external factors,the research of text location in scene images is still challenging.In this paper,a new method about text location in scene images is proposed by combing the advantages of Stroke Width Transform(SWT)and the Maximally Stable Extremal Regions(MSER).The details are as follows:First,take advantage of MSER to segment the image,extract text information from the image,then obtain text candidate region of the image.However,the MSERs of the initial extraction is usually in irregular state,which adds some resistance to the subsequent text location work.Therefore,this paper uses an affine invariant method to synthesize irregular candidate MSERs into ellipses.However,some of the background areas are also treated as text regions after synthesizing.In order to eliminate these nontext areas,this paper has developed some filtering strategies,mainly including: restrictions on the height and width of characters,restrictions on the aspect ratio of characters,restrictions on the edge density of characters.The experimental results show that they can effectively remove obviously non-text areas after a preliminary filter.Then extract the stroke width of each candidate region by SWT.There are many non-text elements appear in the stroke width drawing extracted at the beginning.In order to eliminate these non-text areas,this paper sets out a series of heuristic rules,mainly including:(1)Limit the range of the character aspect ratio to remove some connected regions that are too long or too short.(2)Limit the range of the ratio of the connected regions diameter to the stroke width median.(3)Restrictions on the height of character to prevent some oversized or undersized text areas from being deleted.(4)The threshold is set to half of the stroke width average value to remove some common interference elements such as leaves.In addition,in order to enable characters to be connected into text lines,this paper defines the constraints that the two candidate connected components should satisfy,including median ratio of the stroke width,height ratio,and character space.Finally,we choose some representative images from ICDAR2003 dataset for experimental verification.The results show that the new method of MSER+SWT has achieved a better locating effect(precision up to 76%,recall rate of 61%),and the speed of text locating has been significantly improved.