Image Hashing
Mark Johnson
(Professor Kannan
Ramchandran)
Fannie and John Hertz Foundation
We consider the problem
of mapping an image to a short binary string, known as image hashing. The image
hash function should have the properties that perceptually identical images
should have the same hash value with high probability, while perceptually
different images should have independent hash values. In addition, the hash
function should be secure, so that an attacker cannot predict the hash value of
a known image. An image hash function can be used to search and sort an image
database, or to select frames in a video sequence for watermark embedding.
We construct an image hash function by splitting it into two stages. In
the first step, a feature vector, which should capture the important perceptual
aspects of the image, is extracted. In the second step, the feature vector is
securely compressed. We perform this secure compression by first adding a dither
sequence to the feature vector and then compressing the result with a
distributed source code. We prove that the mutual information between the image
and the hash value goes to zero if the dither is drawn from a suitable
distribution.
Figure 1: Block diagram of image hash
function
Figure
2: Secure compression of feature vector
- [1]
- M. Johnson and K. Ramchandran, "Dither-Based Secure Image Hashing Using
Distributed Coding," Proc. IEEE Int. Conf. Image Processing, (Barcelona,
Spain), September 2003.
Contact: Mark Johnson