Identifier

etd-07132007-091706

Degree

Doctor of Philosophy (PhD)

Department

Engineering Science (Interdepartmental Program)

Document Type

Dissertation

Abstract

Due to the recent improvements in digital photography and storage capacity, storing large amounts of images has been made possible, and efficient means to retrieve images matching a user’s query are needed. Content-based Image Retrieval (CBIR) systems automatically extract image contents based on image features, i.e. color, texture, and shape. Relevance feedback methods are applied to CBIR to integrate users’ perceptions and reduce the gap between high-level image semantics and low-level image features. The precision of a CBIR system in retrieving semantically rich (complex) images is improved in this dissertation work by making advancements in three areas of a CBIR system: input, process, and output. The input of the system includes a mechanism that provides the user with required tools to build and modify her query through feedbacks. Users behavioral in CBIR environments are studied, and a new feedback methodology is presented to efficiently capture users’ image perceptions. The process element includes image learning and retrieval algorithms. A Long-term image retrieval algorithm (LTL), which learns image semantics from prior search results available in the system’s transaction history, is developed using Factor Analysis. Another algorithm, a short-term learner (STL) that captures user’s image perceptions based on image features and user’s feedbacks in the on-going transaction, is developed based on Linear Discriminant Analysis. Then, a mechanism is introduced to integrate these two algorithms to one retrieval procedure. Finally, a retrieval strategy that includes learning and searching phases is defined for arranging images in the output of the system. The developed relevance feedback methodology proved to reduce the effect of human subjectivity in providing feedbacks for complex images. Retrieval algorithms were applied to images with different degrees of complexity. LTL is efficient in extracting the semantics of complex images that have a history in the system. STL is suitable for query and images that can be effectively represented by their image features. Therefore, the performance of the system in retrieving images with visual and conceptual complexities was improved when both algorithms were applied simultaneously. Finally, the strategy of retrieval phases demonstrated promising results when the query complexity increases.

Date

2007

Document Availability at the Time of Submission

Release the entire work immediately for access worldwide.

Committee Chair

Dr. Gerald Knapp

Share

COinS