| To create a visual-data-based image retrieval system, image representation and matching are two major interrelated problems. Image representation requires a procedure to extract features which are rich enough to describe the image characteristics, yet succinct enough to enable an efficient and efficacious matching. Image matching is to identify features between two images so that their resemblance level can be found. This thesis presents the theoretical background and new methods for implementation and analysis of a multimedia image retrieval system, which essentially composes subsystems performing image retrieval using color and shape features respectively.; The frame work of our system is presented, and the models and configuration of each subsystem are discussed. Our focus is especially put on the feature encoding scheme. Several proposed algorithms, such as the CNS-based supervised clustering algorithm for color space quantization, the global-searching multilevel polygon splitting and pair-growing polygon splitting algorithms for feature extraction of an object shape, are described in great details. New concepts, such as Color-Loss, errorbound, etc., which are proven to be effective measurement metrics for optimal selection of color clustering algorithms in creation of a similar color retrieval system, are introduced. Geometric hashing indexing structure is also investigated in our study for shape data management. Substantial experiments have been conducted to analyze and validate the systems over a large set of database images including nature scene color images and a collection of marine animal contour images. Our results have demonstrated the robustness and effectiveness of our system. |
