With the stead development of broadcast TV, the industry accumulates large amount of multimedia data. Due to the lack of effective information of structure and index, it is very difficult to conveniently search from such amount of broadcast multimedia data. On the other hand, because of its specific industry and audiences, after many years of development, the multimedia data of broadcast TV develops its own characteristics. In order to exploit the multimedia data more effectively, it is urgent to construct a management and retrieval system for broadcast TV multimedia data. This thesis covers several bottleneck problems in program layer structure analysis and visual content retrieval, including audio/video template matching, repeated sequence detection, similar image retrieval as well as TV logo recognition. The objective is to research and develop working retrieval system for broadcast TV multimedia data. In particular, the main contributions of this thesis are summarized as follows: This thesis proposes a novel video fingerprinting feature--Global Binary Pattern (GBP). The feature adopts an analysis method combining statistical and structure information. It computes statistical information in local region of image, which preserves robustness of statistical feature, and then computes structure information in global area, which enhance the discriminative power. In the research of audio/video template matching technique, through the deep analysis of characteristics of audio and visual content of broadcast TV data, this thesis proposes a program retrieval framework of fusing audio and visual information. Compared with the method using only single modal feature, proposed framework utilizes the visual feature and the audio feature simultaneously, which improves the accuracy of program video retrieval. This thesis proposes a robust repeated sequence detection method in broadcast streams. The method transforms the repeated sequence detection problem to decoding problem of Hidden Markov Model(HMM). By defining the constraints of state transfer, it relaxes the strict consistency constraints of time. On the basis of it, it defines the detection objective as Maximal Loosely Connected Sequence(MLCS) and adopts a Viterbi-like algorithm to detect them. Due to Viterbi-like algorithm, the proposed method can select the optimal sequence from several redundant state sequences. In addition, we apply the repeated sequence detection to repeated program detection. In order ...
修改评论