Video Clip Retrieval Based On The Spectral Features Of Association Graph

With the development of internet technology and video capture, editing, storage technology, more and more personal video clips are uploaded to the Internet. Faced with a deluge of video data, how to quickly and effectively analyze and search video content become the priority research areas. In this case, more and more research focus on content-based video research. Content-based video clips retrieval can be broken down into shot segmentation, the key frame extraction, the key frame feature extraction, retrieval of similarity video clips. Video clip retrieval based on the spectral features of association graph as a new retrieval system, besides the advantages of content-based retrieval, and constructed the association graph of the video clips, described the video clip with the spectral features of association graph.Shot segmentation and key frame extraction are pre-treatments for video retrieval; it directly affects the final search results. This paper describes the main methods of shot detection and key frame extraction, based on gradient threshold a gradual shot detection method is proposed, Satisfactory experimental results are achieved.The association graph of the video clips was constructed, by analyzing the adjacency matrix of the association graph, the spectral features of adjacency matrix were extracted, which includes the leading eigenvalues, inter-mode adjacency matrices and inter-mode edge-distance. These spectral features are then embedded into the principal component analysis (PCA) and independent component analysis (ICA) model space, for the purpose of cluster analysis based on k-means. Experiments show that the clusters of the spectral features have made very good results under different modes of space. It shows that video clips described by association graph are effective. We applied this cluster method to classify the video library.Study the framework of video clips retrieval based on the spectral features of association graph, and extracted three underlying features from a collection of key frames of each lens in video clips. After that, we constructed three kinds of association graph, and extract a variety of spectral features by calculated the three similar matrix, describe the video clip with the spectral features. A effective classification on the spectral features of the video clip by using k-means clustering, and get the results of video search through the series feature aggregation(SFA).Finally, we summarized the work of the thesis, and points out some future research directions which will be conducted.