||Extracting Hierarchical Structure of Web Video Groups for Realizing Advanced Web Video Retrieval
Due to the widespread use of video hosting services such as YouTube, more and more Web videos, i.e., video materials on theWeb, are being uploaded and accessed by many users. From such a social background, it is necessary to develop a method that enables effective retrieval of desired Web videos. However, there is a limitation that existing search engines require usersto input suitable keywords as a query to accurately retrieve the desired Web videos. To overcome this limitation, this thesis proposes a method for extracting the hierarchical structure of Web video groups on the basis of the relevance between heterogeneous features, i.e., visual, audio and textual features and features of link relationships between Web videos. By utilizing the relevance between heterogeneous features, advanced Web video retrieval be-comes feasible. Specifically, the proposed method calculates latent features on the basis of the correlation between visual, audio and textual features. Then the combination use of the latent features and features of link relationships enables extraction of the hierarchical structure of Web video groups. Furthermore, this thesis proposes a method for improving the proposed method, which enables application of the proposed method to many Web videos. The improved method can efficiently calculate latent features of Web videos on the basis of a clustering scheme. Furthermore, a graph, which can handle many Web videos by a small number of nodes, is constructed and efficient extraction of the hierarchical structure becomes feasible. Moreover, this thesis proposes a method for further improving accuracy and efficiency ofextraction of the hierarchical structure. Specifically, this method utilizes only local features of link relationships and enables accurate and efficient extraction of the hierarchical structure. Experimental results for Web videos collected by using YouTube have confirmed that the proposed methods in this thesis enable retrieval of desired Web videos even if users cannot input suitable keywords as a query.
Hokkaido University（北海道大学）. 博士(情報科学)