Degree Name

Doctor of Philosophy


School of Computer Science and Software Engineering


This thesis investigates three major issues in the active field of content-based image retrieval (CBIR), which are feature aggregation for similarity measure, robust contentbased image retrieval and retrieval model by incorporating background knowledge.

Feature aggregation computes image similarity by fusing multiple distances obtained using pairs of visual feature and distance measure. To explore the mechanism of various aggregation functions, a new feature aggregation framework is proposed to unify multiple existing methods. With multi-example query, each visual feature is of different importance to different query examples, which results in a localized feature aggregation method. A series paradigm is proposed for feature aggregation to exploit the discriminant power of individual visual features, in contrast to conventional parallel paradigm.

The robustness of CBIR systems can affect system performance and user satisfaction. This thesis identifies and addresses two robustness problems in CBIR. Firstly, an unclean query can in uence the retrieval accuracy since some noisy query examples are unable to describe user information need. To address this problem, a robust CBIR scheme is proposed by incorporating noise tolerant classification techniques. Secondly, in large image collections, some classes are prede ned while some are hidden. The CBIR systems employing image classification can not effectively handle the queries associated with the hidden classes. To address this problem, a robust CBIR scheme is proposed.

The retrieval performance can be improved effectively by incorporating background knowledge. For this purpose, a bag of images (BoI) model is proposed for CBIR. The BoI model can express background knowledge more effectively than conventional pairwise constraints, which are demonstrated by two applications. Based on the BoI model, a new clustering method is developed, which can learn a localized similarity measure for each image cluster. A number of experiments validate the effectiveness of the BoI model.