基于概率的跨媒体检索方法分析-计算机软件与理论专业论文.docxVIP

华中科技大学硕士学位论文 华 中 科 技 大 学 硕 士 学 位 论 文 II II Abstract With the development of computer and network technology, the amount of information on the net has shown an explosive momentum, and the type of multimedia varies. Today, users are no longer satisfied with one modality retrieval, which leads to the development of semantic-based cross-media retrieval technology. The paper first clarifies the object of cross-media retrieval including multimodal media interaction retrieval and multimedia document retrieval, and discusses the two main problems in cross-media retrieval including content gab between different modality objects and semantic gab between low-level features and high-level emotion. To solve these problems, dimension reduction, machine learning, et al. are available. Since the heterogeneous features can not be compared, the canonical correlation analysis, which maps two variants into an isomorphic feature space with maximum correlation maintained, is applied to solve the content gab problem. To bridge the semantic gab, a classification model is employed to extract the semantic feature of multimedia objects. When performing the multimodal media interaction retrieval, the retrieval algorithm not only consider the semantic correlation between objects but also the similarity between the objects of the same modality, which is implement by the construction of media correlation graph and the introduction of transition probabilities between the objects of the same modality. When performing the multimedia document retrieval, we fuse the semantic feature of different modality objects into a single one, and the multimedia document retrieval can be simplified as a single modality retrieval problem. Experiments show that the retrieval algorithm based on probability has certain advantages in effectiveness and efficiency. Keywords: Cross-media Retrieval, Canonical Correlation Analysis, Classification, Feature Fusion, Semantic Probability Vector PAGE IV PAGE IV 目录 摘 要