Efficient adaptive retrieval and mining in large multimedia databases

Multimedia data ranging from images to videos and time series is created in
numerous scientific, commercial and home applications. Access to increasingly
large data volumes stored in multimedia databases is a core task to
retrieve similar objects or to generate an overview of the entire content. Examples
include retrieval of similar magnetic resonance images for diagnostic
purposes, or automatic detection of customer segments for sales promotion.
Meaningful retrieval and pattern detection require content-based methods
that describe the relevant characteristics of multimedia objects. As opposed
to manual keyword annotation techniques that are typically infeasible for
large data volumes, content-based approaches use similarity models to process
multimedia data. Similarity models specify appropriate features and
their relationship for effective content based access.
As most multimedia features require many different attributes, high dimensionality
of multimedia features and huge database sizes are major challenges
for efficient and effective retrieval and mining.
In this work, very common feature types for multimedia data are studied:
histogram and time series data. Histograms are used for a variety of
features such as color, shape or texture. Time series data is prevalent for
sensor measurements, stock data, and may even be applied to shapes and
other features as well. For these data types, effective adaptable similarity
3
models are usually computationally far too complex for usage in large high
dimensional multimedia databases. Therefore efficient algorithms for these
effective models are proposed.
In this work, indexing techniques are used that allow for efficient query
processing and mining by restricting the search space to task relevant data.
Multistep filter-and-refine approaches using novel filter functions with quality
guarantees ensure that fast response times are achieved without any loss of
result accuracy.
This thesis is structured as follows: first, in the Preliminaries, an overview
over the thesis and the major challenges in multimedia retrieval and mining
is given. Part I discusses histogram retrieval, Part II studies time series retrieval.
In Part III, efficient and effective histogram data mining is proposed,
and Part IV presents novel time series mining techniques. Finally, this work
is concluded and future research directions are suggested.