Research On Natural Scene Image Text Location Methods And Its Application

Natural scene images are taken by cameras and other photographic equipments fromnatural environments. Text location in natural scene images means to detect and locatecharacters from street signs, license plates, bulletin boards, posters, magazine covers, etc.which appeared in the images. This subject can be applied to fields such as guide blind,navigation, intelligent transportation system, information retrieval. It has wide applicationspace and important research significance. How to extract texts from complicated backgroundimages has become a hot issue.After researching, analyzing and comparing many existing text location methods, thispaper presents some new ideas and methods in some key processes of text location.In terms of image segmentation, studied and improved pulse coupled neural networkwhich has good biological basis on image segmentation. It is used to segment image into threevalues, and well adapt to complicated light conditions of scene images. Multiple parametersof the pulse coupled neural network are automatically chose based on the characteristics ofimage itself. It avoids the problem that traditional method needs to choose parameters bypeople, to which the segmentation results are sensitive.In terms of connecting character connected components as a text area, this paper presentsthe method based on fitting lines among similar components. Map centers of the componentsinto the points of the two-dimensional plane, and define similarity between two componentsby their widths, heights and positions, components can be divided into several sets. Take partof thought from random sample consensus, we fit lines among each sets to isolate text lines.This algorithm has high fitness to different size of characters, and solves the problem todecide which line a point belongs to, when the point has a confused position because of twolines are near and tilted.In terms of feature extraction, by studying the stroke feature of English characters, thispaper describes the “erase” effect that one side edges remove part of pixels of binarycharacters by some certain ways. And based on this, proposes a novel feature named as”tinydisplacement coverage”. Process of extracting this feature is simple, and it has goodperformance on the application of screening non-text items.Explores the way to apply the text location algorithm on mobile terminals, and based on the MVC architecture, text location application program was completed on the Androidoperating system.