Research On Application Of Semantic Boosting And Matrix Factorization For Hashing Based Cross-Modal Retrieval

Multimedia retrieval is a long-term research focus and difficult task in computer vision, and traditional multimedia retrieval methods mostly concentrate on data within single feature modality, such as text retrieval and image retrieval. With the explosive growth of the multimedia data in recent years, how to achieve retrieval among different data from different modalities has become a hot issue in multimedia retrieval. Cross-modal hashing method uses hash functions to encode high-dimensional features of multimedia data into low-dimensional binary hash codes, and preserve the similarity of high-dimensional features. Due to its low storage cost and fast query speed, cross-modal hashing has attracted intensive attention in cross-modal retrieval.In this dissertation, two cross-modal hashing methods, i.e., semantic boosting cross-modal hashing (SBCMH) and supervised matrix factorization hashing (SMFH) are put forward. Comparative experiments on two public datasets demonstrate the effectiveness of the proposed SBCMH and SMFH. The main content is summarized as follows:1. Logistic regression and boosting algorithms are integrated into SBCMH. Firstly, to preserve the semantic similarity, multi-class logistic regression is used to project heterogeneous data into a semantic space, respectively. Secondly, to further narrow the semantic gap between different modalities, a joint boosting framework is used to learn hash functions. Finally, the mapped data representations are transformed into a measurable binary subspace.2. The graph Laplacian regularization and the collective matrix factorization technique are applied to cross-modal hashing. Firstly, the local geometric consistency and the label consistency are presented as the graph Laplacian. Then, the collective matrix factorization technique is used to learn hash functions and learn unified hash codes for different modalities of an instance. Finally, the original features of different data from different modalities are transformed into binary codes with the same length.