From: cme@ellisun.sw.stratus.com (Carl Ellison)
Date: Wed, 6 Oct 93 14:10:20 PDT
To: cypherpunks@toad.com
Subject: Re:  Quantifying similar graphic images (was Re: criminal gif upload)
Message-ID: <9310062109.AA22515@ellisun.sw.stratus.com>
MIME-Version: 1.0
Content-Type: text/plain


>Message-Id: <cggmhoe00VolIKwEcW@andrew.cmu.edu>
>Date: Wed,  6 Oct 1993 16:32:52 -0400 (EDT)
>From: Matthew J Ghio <mg5n+@andrew.cmu.edu>
>Subject: Quantifying similar graphic images (was Re: criminal gif upload)

>"Pat Farrell" <pfarrell@gmu.edu> writes:
>
>> We can prove statistical insignificance of duplication using strong
>> hashing functions. Can we find a way to statistically prove "looks like"
>> on a numerical basis?
>
>Yes.  If you were to take an image and divide it into let's say about 20
>sections horizontally, and 20 sections vertically, and then average the
>intensities of all pixels in each of the 400 rectangles formed, you
>would create a fuzzy low-resolution version of the original picture
>which could be used to compare other pictures


You would have a better chance if you took just the low frequency components
of a 2D Fourier transform of the pictures in question -- perhaps at only
certain frequencies -- to get a vector describing features of the picture.

You'd have to choose your 2D frequencies and build a set of such indicators
and then look to see what distance between two vectors suggests that the
pictures are the same.

You'd want to use only the magnitude of the transform, to remove translation
effects.  You could use a sum around a circle of frequencies to remove
rotation effects.

The low res picture by averaging is easily confused by any translation
or rotation of the image.


 - Carl