PicASHOW: Pictorial Authority Search by Hyperlinks on the Web
Ronny Lempel, Aya Soffer

Jonathan Ledlie (jonathan@eecs)
October 2, 2001

The authors argue that a hybrid of two major text-oriented web page
ranking schemes, PageRank and Kleinberg's algorithm, also apply to
images and could be included in search engines with little
modification.  They merge the probabilities of visiting random pages
from PageRank's link analysis with Kleinberg's idea of hubs (pointers
to authorities) and authorities (high quality containers of
information on a given topic) to form topical image collections IC =
(P,I,L,E), where P is the page set, I is the set of images in P, L is
the page-to-page link set, and E is the page-to-image relation.

They attempt to ameliorate problems areas where pages and
images differ with several heuristics.  In particular, they attempt to
account for the condition where images have been copied instead of
linked to.  They hash the first 1k of each image and then compare this
to other hashes of images, arguing that only if the hashes are the
same are the images the same.  Because they also hash the header
information, I think these hashes are too restrictive and, in fact,
they say that "very rarely" do images have the same signature.  The
problem is that headers contain timestamps and cookie information
which would likely be different even if the images were the same.  A
second heuristic is the elimination of banners and other files they
deem unimportant based on aspect ratios and sizes.  Instead, I think
these should be included but then excluded at the user's discretion
along with other image analysis criteria which they mention in the
future work.

Class discussion:
Comparison between Kleinberg/PageRank and PicASHOW.  Different kinds
of relations.  Here we have Page->Page, and Page->Image relations, a
level of indirection.  What matrix do you use?  They end up suggesting
a linear combination (maybe 1/2, 1/2).

- Fingerprints for Uniqueness
The headers are actually part of the image file.  These are
not HTTP headers.
- Noise and Banner
- StopImages: These are manual; is this reasonable?
- Could this also apply to sound files?