Detection of text captions in compressed domain video

Abstract

In this paper, we propose a new method for detecting text captions in MPEG video stream. It is based on the observation that text captions in video frames typically possess high contrast against the background for them to be visible. The method operates on DCT coefficients in MPEG domain. The main contribution of this work is in developing a binarzied contrast feature domain in which the presence of text in video frames can be highlighted. A weighting function is defined to deduce the probability of a text frame. From the set of representative text frames located, further techniques are developed to outline, segment, and recognise text captions. The techniques have been tested on commercial and news videos from the MPEG7 data set and local TCS news video. The results demonstrate that our approach is effective in detecting and locating text captions in general video.