Efficient and Language Independent News Story Segmentation for Telecast News Videos Anubha Jindal * , Aditya Tiwari † , Hiranmay Ghosh ‡ * TCS-Innovation Labs Delhi 249 D&E, TCS TOWERS, UDYOG VIHAR, PHASE-IV, GURGAON, HARYANA-122016 Email: anubha.jindal@tcs.com † Email: aditya.tiwari@tcs.com ‡ Email: hiranmay.ghosh@tcs.com Abstract—A TV news program comprises a continuous video stream containing a number of news stories, interspersed with commercials and headlines. This paper presents a method to detect the story boundaries and to separate out the stories from the other components and from each other. The method is based on movement of ticker text bands and repitition of ticker texts during different parts of a news program. The method does not use any language processing tool and is independent of language of telecast. It uses some simple features to distinguish news from the advertisements and can be used for large- scale news indexing. We produce some test results on channels telecasting in English and few other Indian languages. Keywords-News analytics, Telecast video, Ticker Text I. I NTRODUCTION Several news monitoring agencies need to monitor a large number of national and international news channels round the clock. The manual monitoring of a large number of channels not only requires a huge effort, but is also error-prone. In this context, we had earlier proposed a framework for automatic analysis and indexing of telecast news programs in multiple languages [10]. In this paper, we provide methods for removing advertisements and for isolating the individual stories from a continous broadcast stream. Our method is independent of language of telecast and uses some simple features that can be detected in near real-time, so that the method can be applied in context of large-scale news indexing srvice. Several methods for advertisement detection and story segmentation in news videos have been reported in the liter- ature. Lienhart [8] proposes an algorithm for advertisement detection by combining feature based approach for pre- filtering and recognition based approach for finding the exact borders. However, this algorithm assumes some specific structure for the commercials, which is not followed world- wide. Hua [12] has presented a learning based advertisement detection approach using a set of six visual features and five audio features. Other features used for advertisement detection includes repetition of shots [9], [7] distinctive acoustic features of commercials [4], appearance of black video frames and silence in audio just before and after the advertisements [5], [8], absence of channel logos during commercials [1] and high-frequency scene changes [12], [8] However, modern video editing techniques used in many news channel falsify these assumptions. Contemporary news presentations often use such techniques for attracting viewer attention and for creating sensation. Use of multiple features makes some of these methods more accurate but also more compute intensive. They cannot be used to process large number of telecast channels in near real-time. Chua et al. [2] has proposed a two level multi modal framework for story segmentation, where the shots are classified using decision tree technique into some predefined categories, which is followed by story segmentation using Hidden Markov Model (HMM). This work was extended in [6], where heuristics has been used to classify a story into news and miscellaneous. This approach allowed the algorithm to scale up to process a large number of videos. Colace et al. [3] has used multi level probabilistic framework using HMM and Bayesian network for segmentation and classification. Use of multiple and complex feature set makes these algorithms slow. Speech or Close Captioned Text (CCT) has been used to distinguish the topic being discussed at different time intervals in [11]. However, CCT is not generally available with TV channels, except for a few countries, where it is mandated by law. Automatic Speech Recognition (ASR) technology is also not available for many languages and wherever available, cannot generally cope up with regional variation of accents. Thus, these algorithms cannot be universally applied on telecast news programs on any TV channel around the world. We propose new methods for commercial detection and story segmentation in this paper, which are language inde- pendent and be performed in near real-time. Our algorithms are based on some news editing characteristics, which are found to be invariant over a large number of international, national and regional news channels. In our approach, the video corresponding to a news program is first processed to identify and remove the commercial segments, based on movement in ticker-text positions between news presentation and the advertisements. We have discovered a common