Video content analysis and TV commercial detection

There is a large and rapidly increasing amount of video data on the Internet and in personal or organizational collections. Fast and accurate video search emerges to become an important issue. The need and main technical challenges for video retrieval are similar to those for the content based image retrieval CBIR problem. Lack of meaningful and comprehensive text annotation means that an approach based on content similarity can be promising; and the differences between an often high-level search intention and the low-level features used in content-based search techniques suggest that content-based video retrieval (CBVR) may also suffer from “semantic gap” issues. One area of content-based video search is automated video matching and recognition. It has emerged in many applications in recent years. This thesis analyses the problem of CBVR from related work in the literature as well as some current work in our team, focusing on the relationship between CBIR and CBVR, open yet well-defined research issues and practical applications of CBVR. To retrieve similar videos to a query clip from a large database, nearest neighbour (NN) search technique for each feature vector is used. Batch nearest neighbour (BNN) search is started as a batch operation that performs a number of individual NN searches. Towards efficient high-dimensional BNN search, a novel approach called dynamic query ordering (DQO) is presented for advanced optimization of both I/O and CPU costs. Computer recognition of TV commercials is one of many interesting areas that have attracted much attention from both the research community and the marketing industry. As the amount of video data is stored in large quantities, it is necessary to propose an efficient technique for video seeking and matching. Automatic real-time recognition of TV commercials is an essential step for TV broadcast monitoring. It comprises of two basic tasks: rapid detection of known commercials that are stored in a database, and accurate recognition of unknown ones that appear for the first time in TV streaming. The existing approaches, however, cannot perform robust commercial detection because they rely highly on the assumption that black frames are inserted before commercial breaks. In this thesis, a real time recognition system – CRS and SOS based approach is proposed to address these challenging issues. The recognition system consists of these components- digital TV capturing, scene breaks detection, fingerprint generation, commercial break and boundary detection, known and unknown commercial recognition. This thesis will describe the details of these components.