On the segmentation of text in videos
A. Wernicke, Rainer W. Lienhart · 2002
A new and robust multi-resolution approach of localizing and segmenting text in videos is proposed. The approach has been tested extensively on a large variety of video frame sizes such 352/spl times/240 up to 1920/spl times/1280 and a large representative set of video sequences such as home videos, newscasts, title sequences and commercials. 95% of the text bounding boxes in videos were localized correctly. 80% of all characters were segmented correctly, while 7.8% characters were damaged. 90% of the correctly segmented characters were recognized correctly by standard OCR software.