Research On Extracting Texts In Nature Scenes Based On Interested Points

Texts in natural scenes contain abundant semantic information. Extractingand identifying these texts is very important for completely understanding thenatural scenes images. However, because of complexity and uncertainty ofnatural scenes, there are many difficulties to hit the mark. Most of the existingmethods are purposeful and lacking of ability to go through various complicatedsituations.After in-depth research about the existing methods of extracting texts fromimages with natural scenes, a new text-extracting method based on clusteringcolors at interested points and spatial arrangement analysis was presented. Themethod divides text-extracting mission into two stages: stage of roughly textlocating and stage of accurately text-extracting.The purpose of roughly text location is to exclude obvious non text areas.It can reduce the range of image in follow-up proceeding. To overcome theproblem of text-discriminate threshold in traditional methods based onedge-density are too sensitive and less adaptability, the new method employed atactic called color distribution density analysis. The tactic used intensity ofcolor changes to locate text area, and combined a global threshold and a localself-adaptive threshold to suppress the background and emphasize texts. Thismake text-discriminate threshold can be selected more easily and have moreadaptability.In the stage of accurately text-extracting, according of two importantcharacters of texts in natural scenes: colors of texts are very different withbackground and spatial arrangement of texts is very regular, the new methoduses clustering colors at interested points to catch the text colors of candidatearea first. Compared with general clustering methods, it avoided the influenceon clustering results caused by background color with huge area, and improvedthe accuracy of text-color catching. After that, new method splited candidateimage into several sub-images so as to divide text-elements from complicatedbackground. The purpose is to guarantee capacity of resisting disturbance offollow-up process. To override the influence caused by changes of font, size,shape on traditional methods based on texts own features, new method provideda set of efficient character representations to describe spatial arrangement oftexts and used them as primary basis to discriminate text or non-text area. Itperformed very well. Experiment show that, text-extracting method based on clustering colors atinterested points and spatial arrangement analysis has strong adaptability indark condition, complex background, changes of text size, color, and shape.