Compressed Domain Techniques to Support Information Retrieval Applications for Broadcast News Videos

Huge amount of video data gets generated every day which needs to be efficiently processed and effectively presented to the users for retrieval of relevant information. The focus of this paper is primarily to explore techniques which can support IR applications on large scale video databases. Almost 25 times faster approximation of a widely used color representation is proposed by using information from the MPEG compressed domain. A scheme is presented for real time matching of videos in online video feeds. LSH based indexing is performed for handling efficient near neighbor queries and support fast clustering. A video clip can be searched in more than 100 hours of video data in few seconds. The techniques are verified on large video data with the help of a specially designed experimental setup. The proposed techniques are shown to be significantly better than several existing and state of the art algorithms.