Context-based similarity retrieval and indexing in large image databases

Date of Award




Degree Name

Doctor of Philosophy (Ph.D.)

First Committee Member

Mansur R. Kabuka, Committee Chair


A framework for retrieving images by spatial similarity (FRISS) in image databases is presented. In this framework, a robust retrieval by spatial similarity (RSS) algorithm is defined as one that recognizes both directional and topological spatial constraints and is able to recognize images even after they undergo translation, scaling, rotation or any arbitrary combination of transformations. An algorithm, $SIM\sb{DTC},$ that satisfies the FRISS specifications is presented. $SIM\sb{DTC}$ introduces the concept of a Rotation Correction Angle (RCA) to align objects in one image spatially closer to matching objects in another image for more accurate similarity assessment. The algorithm was tested using the TESSA image database, an intuition test, and using synthetic images. Analysis shows the robustness of the $SIM\sb{DTC}$ algorithm over current algorithms.To maintain consistency of object names throughout the database, a compact logical shape representation called the Hilbert Morphological Skeleton Transform (HMST) is introduced. The HMST preserves the skeleton properties including information preservation and progressive visualization. An object recognition algorithm, the Hilbert Skeleton Matching Algorithm (HSMA), which renders object similarity as a distance measure is introduced. Results show that the HSMA algorithm achieves a comparable object recognition rate while substantially reducing the complexity of current skeleton matching algorithm.To avoid exhaustive search in large image databases, a multilevel signature file called the Two Signature Multi-Level Signature File (2SMLSF) is introduced. Two types of signatures are generated, one is stored in the leaf of the signature tree and is based on the included domain objects and their spatial relationships. The other is used in the rest of the levels and is based only on the domain objects included in the image. Analytical comparison of the 2SMLSF is given compared to the Two Level Signature File both for storage requirements and search performance. The 2SMLSF significantly reduces the storage requirements. In addition, more general queries involving object existence, spatial relationships, and image variants can be answered using this index structure. Performance of the 2SMLSF indexing method, in most cases, significantly improves over current signature file techniques. A signature generation method independent of object translation, scaling, or rotation is introduced.


Computer Science

Link to Full Text