Managing Large-scale Multimedia Repositories


Capturing and organizing vast volumes of images, such as scientific and medical data, requires new information processing techniques for context of pattern recognition and data mining. In content based retrieval, the main task is the seeking of entries in an image database that are most similar, in some sense, to a given query object. The volume of the data is large, and the feature vectors are, typically, of high dimensionality. In high dimensions, the curse of dimensionality is an issue as the search space grows exponentially with the dimensions. In addition, it is impractical to store all the extracted feature vectors from millions of images in main memory. The time spent accessing the feature vectors on hard storage devices overwhelmingly dominates the time complexity of the search. The time complexity problem is further emphasized when the search is to be performed multiple times in an interactive scenario. One of the main contributions of this dissertation is to enable efficient, effective, and interactive data access. We introduce a modified texture descriptor that has comparable performance but nearly half the dimensionality and less computational expense. Moreover, based on the statistical properties of the texture descriptors, we propose an adaptive for approximate nearest neighbor search indexing approach. In content-based retrieval systems, exact search and retrieval in the feature space is often wasteful. We present an approximate similarity search method for large feature datasets. It improves similarity retrieval efficiency without compromising on the retrieval quality. We also address the computation bottleneck of a real-life system interface. We propose a similarity search scheme that exploits correlations between two consecutive nearest neighbor sets and considerably accelerates interactive search, particularly in the context of relevance feedback mechanisms that support distance metric update approach. In multimedia query processing, the main task is the seeking of entries in a multimedia database that are most similar to a given query object. Since feature descriptors approximately capture information contained in images, they often do not capture visual concepts contained in those images. Semantic analysis of multimedia content is needed. We introduce a framework for learning and summarizing basic semantic concepts in scientific datasets. Moreover, we present a method to detect coarse spatial patterns and visual concepts in image and video datasets. Experiments on a large set of aerial images and video data are presented.
[PDF] [BibTex]
Jelena Tešic,
Ph.D. Thesis, University of California, Santa Barbara, Sep. 2004.
Node ID: 402 , DB ID: 203 , Lab: VRL , Target: Thesis
Subject: [Managing Multimedia Databases] « Look up more