Possible Improvements
Here are just a few of the possible improvements I've thought about
for Eikon, in no particular order. My favorites include support
for intra-image searching, using the filenames of images to increase
the relevance of the search, and relevancy scores for results.
- Eikon could form an integral part of an application to allow a
user
to search, sort, and categorize all the images on his computer
or corporate intranet.
- A huge internet-wide image search engine would be amazing.
- It might be nice to store an image thumbnail to show the user
(a la Google's page caching) in case a site is down, or the image
is just too huge.
- It would be cool to be able to get a result back, and allow the
user to then say "Give me all the images from this domain."
- It would probably be a good improvement to save the URL of the
image's webpage as an optional attribute, so that
originating page can be linked to in search results.
- Intra-image searching. There has been some
research on this, but
I have not studied it extensively.
- Since URLs are also unique (assuming that it doesn't point to
a program which generates JPEGs ;-), it would save a lot of time
if I looked to see if that URL has already been retrieved
instead of downloading it again. Downside: It will take more work to
make sure the database stays fresh.
- Display percent match ranking statistics with search results.
- A visualization tool showing how many and where
matches occurred.
This would be great for debugging, and it would also be a neat
thing for the user to look at.
- Experiment with other wavelet basises.
- Experiment with other scoring weights instead of just using
the numbers in the research paper.
- Experiment with the number of coeffients stored.
Finkelstein,
et al.
used 40 for scanned queries and 60 for drawn queries; I have used
60 coeffients.
- Determine if there is a cutoff score where images are
clearly not related, and stop returning results at that
point.
- Hybrid scoring, with multiple scoring metrics, like the
WHASSUP project. Extensions on this idea include using multiple
color spaces and wavelet basises for each image. Images that match
will match across multiple wavelet basises and color spaces, but
the images which do not match are different.
- Partitioning of the database data across multiple
servers or a P2P network.
- Search a P2P network for similar images by sending out an
Query objects (serialized as XML) to other nodes and asking for
similar images.
- Use Eikon as part of a comprehensive image metadata archive,
along with other information uploaded by users or professionals.
Eikon could be used as part of the search functionality of such
an archive, to help properly tag images with other metadata.
Luke Francl
Last modified: Mon Mar 4 22:00:18 CST 2002