Localization of caption texts in natural scenes using a wavelet transformation

Abstract

Automatic extraction of text from multimedia contents is an important problem that needs to be solved in order to obtain more effective retrieval engines. Recently, Crandall, Antani and Kasturi have shown that a direct analysis of certain DCT coefficients can be used to locate potential regions of caption text in MPEG-1 videos. In this paper, we extend their proposal to wavelet-coded images, and show that localization of text superimposed in natural scenes can also be effectively and efficiently performed by a wavelet transformation of the image followed by an analysis of the distribution of second order statistics on high frequency wavelet bands. © Springer-Verlag 2004.