Orientation-independent headlines area extraction by using a half-cosine function wavelet network

This paper describes human-like headlines area extraction from a real-world image by using a half-cosine function wavelet network. The method exploits only a property of headline's appearance that a line width of headline's character is comparatively wide. The headline's area can be picked up by reconstructing with selected basis functions corresponding to headlines. In this paper, headline's extraction from the realistic image, which contains various orientations' headlines written by several languages, is examined and the validity of the proposed method is discussed.