Automatic text detection and removal in video sequences

This paper proposes an approach for automatic text detection and removal in video sequences based on support vector machines (SVMs) and spatiotemporal restoration. Given two consecutive frames, first, text regions in the current frame are detected by an SVM-based texture classifier. Second, two stages are performed for the restoration of the regions occluded by the detected text regions: temporal restoration in consecutive frames and spatial restoration in the current frame. Utilizing text motion and background difference, an input video sequence is classified and a different temporal restoration scheme is applied to the sequence. Such a combination of temporal restoration and spatial restoration shows great potential for automatic detection and removal of objects of interest in various kinds of video sequences, and is applicable to many applications such as translation of captions and replacement of indirect advertisements in videos. (C) 2003 Elsevier B.V. All rights reserved.