| While significant progress has been made in content-based image retrieval over the last several years, there has been less work that has addressed issues related to overall system design from a networked system viewpoint. Since most of the image retrieval services are requested by remote users, possibly mobile users with limited device resources, the network environments of the CBIR systems will affect the overall performance of the image retrieval process. Currently, this process is optimized and well-tuned on stand-alone workstations with traditional performance metrics such as recall and precision. These metrics do not guarantee a satisfactory user experience in a general network scenario.; In this dissertation, we investigate how to enhance semantic relevancy in the retrieval process by semantic feature extraction and relevance feedback. In addition, we propose prefetching and scalable image delivery (progressive and region of interest based delivery of images) to reduce network latency in the retrieval process and to adapt the retrieval process to the network bandwidth and the capabilities of the user device (especially, size of the display screen). However, these two goals, the maximization of semantic relevancy and the minimization of retrieval latency, sometimes conflict with each other. If the user's resources are limited, a slight sacrifice of semantic relevancy can improve the overall performance of image retrieval. As such, we investigate the issues on the joint optimization of these two goals as a specific goal of this dissertation. |
