Information Technology Reference
In-Depth Information
based on addition, such as quantization watermarking [17,18], different energy wa-
termarking (DEW) [19] and non-additive watermarking [20], have emerged.
The reason that embedding capacity of schemes using addition of PN sequences is
low is that the information carrying capability of each sequence has not been fully
exploited. In this regard, we propose an improved multi-bit scheme using a set of
orthogonal binary sequences, leading to significant improvements in performance.
This technique can be incorporated into various public-watermarking frameworks
based upon addition of binary sequences and developed for various digital media
including images and audio signals.
The paper is organized as follows. Section II analytically discusses performances
of multi-bit watermarking schemes based on addition of PN sequences and using a
one-bit-per-sequence strategy. Section III proposes a novel scheme using orthogonal
sequences to achieve increased capacity. In Section IV, performance of the new
method is studied, and simulation results presented. Section V concludes the paper.
2
Multi-bit Watermarking Based
on One-Bit-per-Sequence Scheme
Multi-bit watermarking using a one-bit-per-sequence scheme is a straightforward
extension of the single-bit methods. A single-bit method adds a PN sequence as a
watermark into host data. Cross-correlation between the received data and the known
watermark is computed in detection. If the correlation is greater than a predefined
threshold, the received data is judged as
marked
, otherwise
not-marked
.
Suppose that
I
is the host data.
S
is a PN sequence of length
N
whose elements are
either +1 or
−
1. The embedding scheme can be expressed as:
I
'
j
)
=
I
(
j
)
+
α
S
(
j
),
j
=
0
1
,
N
−
1
,
(1)
where
I'
is the watermarked signal, and
the strength of the mark. Modifying the
watermark amplitude according to the host data, or introducing characteristics of the
human perseptual system,
H
, to improve imperceptibility, Equation (1) is modified as
α
,
I
'
j
)
=
I
(
j
)
+
α
H
(
j
)
S
(
j
),
j
=
0
1
,
N
−
1
(2)
where
H
is related to perceptual models and may be a function of frequencies and
spatial/temporal properties. In the presence of additive interference (channel noise or
hostile attack), the received signal becomes
.
I
'
j
)
=
I
'
j
)
+
N
(
j
)
(3)
In this study, the host data are pre-processed in some way before embedding in or-
der to achieve spectrum equalization, that is, to make each coefficient possess a uni-
form energy statistically. For example, shuffling the host data pseudo-randomly prior
to transform can remove correlation between adjacent samples [21,22]. Therefore, the
coefficients in the transform domain are i.i.d. Gaussian with a zero mean and an iden-
tical standard deviation
I
. With spectrum equalization, Equation (1), instead of (2),
can be used to simplify analysis, and make it possible to directly study the relation
