Jun Tian talks with
ScienceWatch.com and answers a few questions about
this month's Fast Moving Front in the field of Computer
Science.
Article: Reversible data embedding using a
difference expansion
Authors: Tian,
J
Addresses: IEEE TRANS CIRC SYST VIDEO T, 13 (8): 890-896
AUG 2003
ICompress Technol, Tualatin, OR 97062 USA.
ICompress Technol, Tualatin, OR 97062 USA.
Why do you think your paper is highly
cited?
Reversible data embedding is a difficult problem. It embeds—or
hides—information into a digital content without degrading its
perceptual appearance, and has the capability to restore the content to its
original, pristine state, bit by bit and exactly. The reversibility is the
challenging part. Before our work, there had been a number of exciting
breakthroughs in this area, yet the embedding capacity (i.e., the maximal
amount of information that could be embedded) was rather limited.
We approached the problem from the different direction of utilizing integer
transform, presenting a simple yet effective concept of difference
expansion for data embedding, and employed the transform
invariants—which is an invariant set of the embedding
process—to achieve reversibility. It increased the embedding capacity
four-fold (from 0.5 bpp to 2.0 bpp for an 8 bpp grayscale image), and the
perceptual quality, after data embedding, was also among the best.
Does it describe a new discovery, methodology, or
synthesis of knowledge?
It describes a new method for reversible data embedding. Previously,
reversible data embedding methods were mostly compression-based: one
selects a region or feature of a digital content, losslessly compresses it,
and the space saved from compression will be used for data embedding.
"Reversible data embedding has a big
potential in sensitive imagery, such as
military data, medical data, and digital
reproduction of fine
arts."
Our method applies an integer transform (for example, integer wavelet
transform), and discovers extra storage space in high frequency domains
(which are usually small magnitude coefficients). The data embedding is
achieved by a simple bit-shift process, which is called a "difference
expansion."
Would you summarize the significance of your paper
in layman's terms?
Our paper developed a new diagram for reversible data embedding, i.e.,
integer transform + difference expansion + location information. The
integer transform inherits the advantage of traditional decorrelation
transforms, with the additional benefit of reversibility.
The difference expansion technique embeds (and extracts) information by bit
shift operation, which gives a large embedding capacity, and maintains the
perceptual quality as well. Perceptually, difference expansion is similar
to mild sharpening in the mid tone regions.
The location information keeps record of where difference expansion
happens, and will be part of the information to be embedded, such that the
decoder could extract the embedded information and restore the original
content, without referring to any side information.
How did you become involved in this research and
were any particular problems encountered along the way?
Reversible data embedding problems first appeared in late 90s. Inspired by
the groundbreaking work of Dr. Chris Honsinger et al., of the
Eastman Kodak Company, and also pioneering work from Prof. Jessica
Fridrich's group at SUNY Binghamton.
I was excited by this fast-growing area of research and devoted my efforts
to investigating new methods. I received tremendous support from my
manager, Dr. Steve Decker, to develop high capacity and high perceptual
quality reversible data embedding. Lots of failure, and lots of
frustrations happened during the development; yet when we finally got
there, it was well worth the effort.
Where do you see your research leading in the
future?
Reversible data embedding has huge potential uses in sensitive imagery,
such as military data, medical data, and digital reproduction of fine arts.
In these scenarios, every bit of information is important and should be
preserved. With reversible data embedding, for example, metadata could be
embedded in the content, which eliminates the requirement of additional
storage and is more convenient to manipulate.
Additionally, reversible data embedding is a natural content authentication
tool, where the content message authentication code (MAC) could be embedded
for self-authentication. Essentially, reversible data embedding provides an
additional information channel, which coresides with the digital content.
Do you foresee any social or political implications
for your research?
For engineering, especially information communication, yes, I can foresee
implications.
Jun Tian, Ph.D.
Senior Researcher
Core Network Research Department
Huawei Technologies (USA)
Bridgewater, NJ, US