| In recent years, with the development of technology on network and digital media, more and more multimedia information which include not only text and audio information but also video and image can be easily used by people. Meanwhile, traditional information retrieval method can not deal well with these multimedia contents. So people hope to build a new framework to process and analyze these multimedia data effectively, especially for video clips which combined text, image and audio and had a stronger representation than other media types.Based on these requirements, methods of content-based video information retrieval were proposed. The video contents are described by low-level features such as color, texture, shape, motion and audio. Video information retrieval is implemented by computing the similarity of features of the video clips.Nowadays research work on content-based video information retrieval has made great progress. It includes many research domains such as feature extraction, video structure analysis, video summarization, key-frame selection, database management, user interface, etc. However, due to some limitations such as complexity of video data and the semantic gap between low-level feature and high-level semantics, it still can not meet users' requirement. There still exist many challenges in this domain.The research work in this thesis mainly focus on video structure analysis which is proposed to solve the problem that video itself does not have a nature structure and gives a hierarchical understanding for video content. Generally, video has a structure of frame, shot, scene, and video from the low level to the high level representation. Frame is the basic unit of video. Shot is an unbroken sequence of frames from one camera. Shot is the basis of vide structure analysis. Much research work focuses on detecting shot changes. These methods are effective. However, shot itself can not convey enough semantics information, and one hour video may contain hundreds of shots. Thus it is difficult to analyze video content just based on detected shots. Scene is a group of consecutive shots that focus on an object or objects of interest. One scene can denote a semantic unit. So the effectiveness of scene segmentation is important for the further analysis and understanding of the video content. Also, the results are useful for applications of movie summarization, program segmentation, etc. We focus on scene segmentation in this thesis. Based on the research in this field, we discuss two problems. The first is that most research adopted the idea of clustering. However, how to automatically determine the cluster number still remains a problem. Second, many methods were much dependent on the specified grammar which was used to constitute videos. Thus, we proposed two methods for scene segmentation. Specifically, in the first method, we first use the manifold to explore the essential feature of video structure. Then MCMC (Markov chain Monte Carlo) algorithm is adopted to segment the above features. In the second method, we adopt the idea of JSEG to exploit the coherent characteristic of scene structure based on visual similarity and temporal relations of shots. Then spectral method is used to group shots into scenes. For different videos, both methods can adaptively reflect the structure characteristic that guarantee the effectiveness to various videos by avoiding the subjective affections. Due to the employ of MCMC and spectral method, we can achieve scene segmentation automatically. We tested our methods on different kinds of videos including movie, sports and cartoon, etc. The results demonstrate the effectiveness of our methods.In this thesis, we also introduce the idea of structure analysis into the field of VCR support for P2P VoD systems. Using results of video segmentation results, we not only change dispatched contents but also provide the hierarchal view of video contents. This can greatly improve users' experience. Also, it can alleviate the burden of the system effectively. |
PDF Full Text Request