A Framework for Reliable Text-Based Indexing of Video

Abstract: In this paper we describe our recent research efforts towards reliable and automatic generation of indices for use in content understanding of video. Following our earlier research in temporal shot segmentation of video, we have developed a comprehensive system framework for segmenting an unconstrained variety of text from general purpose broadcast video. In addition, the framework also contains a novel tracking and a binarization algorithm. Also developed to be a part of the above framework are other modules, viz. a novel scene text segmentation method, and a novel text segmentation method which extracts uniform colored text from still video frames. We have thoroughly evaluated the methods which form a part of our framework against a fairly large dataset. The framework applies a battery of methods for reliable localization and extraction of text regions. Towards this, we have developed methods for fusing the results from different methods. More recently, we have extended our interest to localizing and extracting stylized text from video and determining the lifetimes of the video text events. Results from the above research are presented in this paper.