![](bg2.png)
taken into account when embedding the watermark
and, as a result, the maximum-possible
imperceptibility and robustness of the embedded
watermark cannot be guaranteed. In (Nasir et al.,
2008), the watermark was embedded into DC
coefficients of gray-scale images without taking into
account the content of the image.
Based on the fact that the magnitude of DC
coefficients is much larger than any AC coefficients,
DC coefficients can provide more perceptual
capacity than AC coefficients. On the other hand,
DC coefficients are less affected than any AC
coefficients when the watermarked is attacked by
JPEG compression, lowpass filtering and
subsampling operations. Therefore, DC coefficients
are suitable for embedding watermark (Huang et al.,
2005
). Motivated by those observations, in this
paper, we propose a new adaptive and blind image
watermarking method, which is based on the
principle of embedding a watermark in the DC
coefficients of subimages in the DCT domain. These
subimages are obtained through subsampling the
original luminance component Y or the blue
component B of color images in the YIQ or the
RGB color models respectively. In comparison with
existing reported work, our proposed method
possesses significant advantages, which can be
highlighted as: (i) The watermark is embedded in
the DC coefficients to provide more robustness than
using AC coefficients; (ii) The watermark is
extracted without knowledge of the original non-
watermarked image; (iii) The watermark is extracted
directly in the spatial domain rather than applying
the DCT again to the watermarked image; (iv) The
strength of the watermark is determined adaptively
to the contents of the host image to guarantee the
best possible perceptibility and robustness of the
embedded watermark; (v) the DCT and its inverse
are applied only to the selected blocks, which are
used to embed the watermark.
The rest of this paper is structured as follows.
Section 2 describes the adaptive determination of
watermarking strength; Section 3 and 4 present the
embedding and the extraction processes and Section
5 presents some experimental results. Conclusions
are drawn in Section 6.
2 ADAPTIVE DETERMINATION
OF WATERMARKING
STRENTH
A major challenge in designing a watermarking
algorithm is to find a strategy that satisfies the
conflicting objectives that, on one hand, the added
watermark is imperceptible to the human eyes but,
on the other, it should be robust to removal attacks.
The best way to achieve better trade-offs between
imperceptibility and robustness requirements is to
take the characteristics of the non-watermarked
image into account when embedding the watermark.
Chang et al (2005) proposed a technique for
extracting 5 edge patterns directly in the DCT
domain, and proved that DCT blocks of size 8
8
can be classified as certain type of edge patterns.
Jiang et al. (2008) suggested that three edge patterns
rather than five are sufficient to describe and
characterize the visual content of the image in the
DCT domain. Therefore, the proposed method
exploits this block classification scheme to analyze
the visual content and hence determine the
watermark embedding strength.
Via exploitation of Jiang’s classification scheme,
all DCT blocks can be further analyzed as smooth or
non-smooth based on the specific values of the two
DCT coefficients. Non-smooth blocks are then
further classified to determine if they contain both
vertical and horizontal edges or contain one of the
edge patterns. To determine the embedding strength,
we proposed the following adaptive scheme
)
()
⎪
⎩
⎪
⎨
⎧
α
λ≥δδα
λ<δδα
=α
π
π
otherwise
,minifelse
,maxif
edge
22/0texture
12/0smooth
(1)
where α stands for the embedding strength and,
)0,1(
0
X=
δ
,
)1,0(
2/
X=
π
δ
are the absolute value
of the DCT coefficients X(1,0) and X(0,1),
respectively,
1
and
2
are thresholds. The
derivation of
0
,
2/π
does not require any addition
or multiplication and only two DCT coefficients are
used to classify DCT blocks.
Figure 1 demonstrates the classification results
obtained by applying the proposed method to images
with different textures. As an example, Lena
includes large smooth areas with sharp edges;
Peppers includes large smooth areas without sharp
edges and Baboon includes textured areas.
The white areas shown in Figure 1 are classified
as smooth, and thus any small change incurred by
watermarking could be visible. As a result, the
corresponding watermark should have a low
embedding strength. Similarly, the black areas in
Figure 1 are classified as edge or textured blocks,
and hence changes incurred by watermarking would
be less visible. Therefore, the watermark embedding
SIGMAP 2010 - International Conference on Signal Processing and Multimedia Applications
172