Automatic text regions location in video frames

Content-based information retrieval from digital video databases and media archives is a challenging problem and is rapidly gaining widespread research and commercial interest. For a reliable retrieval and intelligent access to video programs, indexing should provide semantic descriptors. One way to include more semantic knowledge into the indexing process is to use the text embedded within images and video sequences programs such as credit titles, ellipse, etc. Text in video is rich in information and easy to use, e.g. by key word based queries. In this paper we propose an automatic text regions location technique in digital video frames. The detected text boxes can then be passed to standard commercial OCR software to obtain the full texts used in the video indexing purpose. Our method makes use of four main techniques in image processing, that is an adaptive binarization, multi-resolution, histogram segmentation and morphologic operations to locate text regions. A new technique for histogram segmentation based on Optimum thresholding is then proposed. The quality of localized text is improved by experimental results that we have driven on a large sample of video frames selected from various kinds of video programs (commercials, TV news, full-length films, etc.). Finally, the results of text regions localization are presented. This work is subscribed among the CMCU Project undertaken