Automatic Recognition And Detection For TV Commercial

The rapid development of network and computer technology results in the explosive growth of the volume of multimedia, particular for the digital videos. To explore the large quantity of information embedded in these videos, people pay much focus on how to efficiently manage, index and retrieve them. As one of the most important parts of digital videos, video commercial has significantly affected and changed people's life. Meanwhile, as one of the most effective media form, commercial has played an irreplaceable role in conveying the product or service information to the consumers. Owing to the growing number and the diversity of commercials, it is desirable to design a smart commercial monitoring scheme to automatically manage and analyze commercials.To alleviate these increasing requirements, this thesis presents an automatic commercial recognition and detection scheme based on the in-depth analysis on commercial characteristics. The following points highlight the two main contributions of this thesis:(1) Pruned Multi-level Successive Elimination for TV Commercial RecognitionIn this thesis, an efficient duplicate matching algorithm, called pruned multi-level successive elimination (PMSE), is proposed for TV commercial recognition. To enhance the efficiency of filtering out the irrelevant candidates from a sizable database, a felicitous pruning strategy is adapted to the multi-level successive elimination by exploiting the similarity relations of all candidates that can be constructed off-line. By progressively partitioning the signatures into finer granularity representation, more candidates can be eliminated with low computational complexity through pruning process at coarse granularity level. Moreover, a well-designed commercial content signature based on visual spatial correlation and LBP-like coding method, is presented to robustly resist the visual perception distortion.(2) TV commercial Detection Based on Explicitly Sharing SubspaceTo alleviate the disadvantage of commercial recognition, i.e. the dependency of the commercial database, an automatic commercial detection scheme based on explicitly-sharing-subspace is proposed in this thesis by exploring the essential difference between the commercials and general programs. Given the two modalities (i.e. visual and audio) of the broadcast videos, an explicitly-sharing-subspace approach is utilized to explore the unified semantic description of these two modalities for the purpose of reinforcing the distinguishing ability of commercials versus general programs. Then, based on this unified description, we apply the SVM classifier to form a final decision whether the shot belongs to commercial or not. In addition, a time continuance based post processing method is trigged out to further alleviate the false alarms existing in the classification process.