Content-based information retrieval (CBIR) has attracted significant interest in the past few years. When given a search query, the search engine will compare the query with all the stored information in the database through nearest neighbor search. Finally, the system will return the most similar items. We contribute to the CBIR research the following: firstly, Distance Metric Learning (DML) is studied to improve retrieval accuracy of nearest neighbor search. Additionally, Hash Function Learning (HFL) is considered to accelerate the retrieval process. On one hand, a new local metric learning framework is proposed - Reduced-Rank Local Metric Learning (R2LML). By considering a conical combination of Mahalanobis metrics, the proposed method is able to better capture information like data's similarity and location. A regularization to suppress the noise and avoid over-fitting is also incorporated into the formulation. Based on the different methods to infer the weights for the local metric, we considered two frameworks: Transductive Reduced-Rank Local Metric Learning (T-R2LML), which utilizes transductive learning, while Efficient Reduced-Rank Local Metric Learning (E-R2LML)employs a simpler and faster approximated method. Besides, we study the convergence property of the proposed block coordinate descent algorithms for both our frameworks. The extensive experiments show the superiority of our approaches. On the other hand, *Supervised Hash Learning (*SHL), which could be used in supervised, semi-supervised and unsupervised learning scenarios, was proposed in the dissertation. By considering several codewords which could be learned from the data, the proposed method naturally derives to several Support Vector Machine (SVM) problems. After providing an efficient training algorithm, we also study the theoretical generalization bound of the new hashing framework. In the final experiments, *SHL outperforms many other popular hash function learning methods. Additionally, in order to cope with large data sets, we also conducted experiments running on big data using a parallel computing software package, namely LIBSKYLARK.
Doctor of Philosophy (Ph.D.)
College of Engineering and Computer Science
Electrical Engineering and Computer Engineering
Length of Campus-only Access
Doctoral Dissertation (Open Access)
Huang, Yinjie, "Content-based Information Retrieval via Nearest Neighbor Search" (2016). Electronic Theses and Dissertations. 5085.