Start the demo
here
Introduction
Due to the ever increasing amount of image data available there is a
need for new, sophisticated tools to search these data. Compared to
text retrieval it is much more difficult to analyze the semantics of
an image (for very restricted environments it is possible, but it
remains unsolved for arbitrary image data). Therefore most image
retrieval systems concentrate on features that describe images on a
syntactical level. E.g., color histograms characterize an image in
terms of the frequencies of different colors.
Approach
Our approach is based on invariant features, i.e. features that do not
vary if the image is transformed by some transformation group (we will
consider translation and rotation here). Schulz-Mirbach introduced an
algorithm for the construction of invariant features
[Schulz-Mirbach:1995] which is very suitable because of its robustness
to slight topological deformations and even to independent motion of
objects within the image. The major advantage is that it does not
require the extraction of objects (segmentation), or distinct points
(key-points) from the image, but can be applied directly to the
original image data.
However, in order to improve the algorithm's robustness in an image retrieval
application - especially for supporting partial matches - we had to
modify it, so that more local information is preserved in the final
features. Thus we constructed feature histograms [Siggelkow,
Burkhardt:1998], which are very similar to the well known color
histograms but consider features drawn from a local neighborhood of
each pixel instead of just using the color value of each pixel
only. Thus we incorporate also textural information.
Recently the method was further enhanced by a fast estimation of the
features instead of a tedious calculation. Thus the extracted features
will have a small error which, however, can
be well estimated [Siggelkow, Schael:1999].
For further information you may want to take a look at my publications
or contact me.
Technical details
Dpending on the configuration, the database server runs on a Linux Debian server, 5 clients are allowed to connect
simultaneously. The clients can be virtually installed on any machine
in the Internet. This means - provided you installed the client
software on your machine - your query image could be analyzed locally
at your site and only the extracted features would have to be
transferred over the Internet, not the image itself (e.g., you might
have a copyright on it and therefore don't want to send it).
The sources are written in ANSI C++ (at least gcc-2.95.2 -ansi
-pedantic doesn't complain;-) and have been successfully
compiled on AIX (IBM), IRIX (SGI), Linux (PC), and even Windows NT (PC) via
cygwin.
The following third party software has been used:
- cvs: version control software
- egcs/gcc: C++ compiler
- libjpeg: JPEG
library
- netcdf:
Unidata NetCDF-File format for storing feature
data. Unfortunately the sources contained a constant defining
the maximum number of variables and attributes per file, this
had to be set higher
- tcl/tk:
for the user interface, 8.0 versions required due to some
changes in later versions of tcl/tk
Author: Sven
Siggelkow
