With the development of technologies, the popularity of the digit mobile devices and social media, multimedia content explosively increases. Multimedia presents rich and vivid content in the form of image, audio and video besides text. Consequently, how to effectively analyze and understand these massive multimedia information with rich contents and forms becomes a hot research topic studying recently. There are increasingly more kinds of features of multimedia data and the dimensionality of features is becoming increasingly high. However, not all the features are helpful to the performance. Most of them are often correlated or redundant to each other, and sometimes noisy. It can promote multimedia analysis and understanding to uncover the data representations reflecting the intrinsic properties from these features. On the other hand, multimedia data are widely described by low-level features and there exists the well-known "semantic gap" between the high-level semantic meaning and the low-level visual features. Fully exploiting the rich context information can learn a compact representation for multimedia data, bridge low-level features and high-level semantics and effectively improve multimedia content analysis and understanding. In addition, the information of multiple media usually have compatibility and complementary characteristics to represent the semantic information. Jointly mining and fusing these heterogeneous information enables to make multimedia analysis and understanding better. For the above issues, based on subspace learning, this thesis makes a study of theoretical researches (feature selection and semantic mapping) and applications (personalized tag recommendation and news retrieval). Our main contributions are summarized as follows. 1. Nonnegative spectral clustering and structural learning-based unsupervised feature selection. To handle the high-dimensional, noisy and redundant features, we propose an unsupervised feature selection framework to jointly exploit the nonnegative spectral clustering and the latent structured analysis. In the feature selection procedure, we propose a novel nonnegative spectral clustering algorithm to learn the label indicate function, which can provide discriminative information for feature selection. On the other hand, we propose to uncover a latent shared structure to mine the feature correlation and assume that the latent structure is a low dimensional linear subspace. The proposed method can effect...
修改评论