01-11-2016, 11:09 AM
1462857707-pixel.pdf (Size: 562.84 KB / Downloads: 1)
ABSTRACT
Robust image watermarking systems are required to be resistant to
geometric attacks in addition to common image processing tasks,
such as JPEG compression. However, robustness against geometric
attacks, such as rotation, scaling and translation, still remains one
of the most challenging research topics in image watermarking. We
propose a new pixel-based watermarking system in which a binary
logo is embedded, a bit per pixel, in the pixel domain of an image.
The encoder of the proposed system is based on a sliding window
embedding scheme that applies the local average quantization index
modulation (QIM), to achieve geometric attack robustness. The decoder
employs a maximum a posteriori (MAP) estimation supported
by Markov Random Field (MRF) model to achieve robust decoding.
Additionally, we demonstrate that the proposed scheme is also robust
against possible watermark removal due to JPEG compression.
1. INTRODUCTION
For digital image watermarking systems, geometric attacks, such
as rotation, scaling and translation, do not distort or remove the
embedded watermark, but instead geometrically and globally
modify the watermarked image to make the watermark decoder
(or detector) unable to re-synchronize the received image. Most
existing robust watermarking systems are block based and/or rely
on the correct synchronization of the image to extract the embedded
watermark. Geometric attacks destroy the synchronization and
render the extracted embedded watermark incorrect or the extraction
process impossible, thus making the watermark undetectable.
Robust image watermarking systems, which are used to address
security concerns, such as copyright protection or copy control,
should guarantee resistance to geometric attacks. Several systems
have been proposed to address the problem of robustness against
geometric attacks[1]. Exhaustive search techniques try all possible
combinations of the geometric distortion and can be computationally
costly or infeasible. Methods that embed a reference pattern
into an image in addition to the robust watermark for aligning
the received image at the decoder can impair either the fidelity or
general robustness of the system. Invariant watermarking systems
are designed to be robust only against certain geometric attacks
and the approaches of autocorrelation and implicit synchronization
also suffer from variant problems. Therefore, robustness against
geometric attacks still remains one of the difficult challenges in
image watermarking research [1, 2].
We suggest that pixel-based watermarking is ideal for addressing
the problem of geometric attacks. The watermark embedding takes
place on each individual pixel - in a binary system, a single bit is
embedded in each pixel. Since the embedding unit is a pixel, the smallest block possible in images, the watermark extraction (decoding)
is not affected by geometric attacks because re-synchronization
problem is obviated. However, we now face another problem,
namely, pixel-based watermarking is usually very fragile even to
common image processing. For example, an attacker can use high
quality JPEG compression to compress the image and remove the
embedded watermark. To make pixel-based watermarking systems
practically applicable, robustness against common image processing
tasks must be improved.
In this paper, we propose a novel pixel-based watermarking system
that is robust against both JPEG image compression and geometric
attacks. For a given image, a binary logo, with the same size
as the host image, is embedded. At each pixel we embed the corresponding
bit of the logo and this embedding scheme makes the
embedded logo resilient to geometric attacks. A sliding window
with the predefined shape visits each pixel and the embedding is
achieved by utilizing the quantization index modulation (QIM) [3, 4]
and exploiting the idea from the local average QIM [5]. The applied
embedding technique makes the logo robust against the JPEG compression.
The possibly distorted watermark (the binary logo) is extracted
and recovered by a MAP decoder (The original image is not
known to the decoder since QIM as a host signal interference rejecting
system does not need the original signal in the decoding process
). Compared with the default Minimum Distance decoder of QIM,
MAP decoder greatly reduces the error-decoding rate and improves
quality of the decoded watermark. The proposed system can be used
to address copyright protection of images where the distinctive logo
can be used to identify the owner or establish claim of ownership.
2. THE PROPOSED SYSTEM
The proposed system consists of a sliding window based local average
QIM (SW-LAQIM) embedding component and a maximum a
posterior probability (MAP) decoder.
2.1. SW-LAQIM embedding
For a given image I with width w and height h, we select a registered
binary logo Z of the same size as I; Z may be constructed through
tiling, scaling or cropping from a logo that is smaller or larger than I.
Assume that s ≡ (i, j), 0 ≤ i<w and 0 ≤ j<h, represents
a site in the image I; Is and Zs then represent respectively the
corresponding pixel in I and the bit in the logo Z. The process of
embedding Z into the host image I is as follows.
For pixel Is, we choose a local window centered at s, as shown
in Fig.1. Let Gs be the set of pixels in the local window, Gs ∈ I.
Let µs represent the average value of the pixels in Gs.
MAP decoding
2.2.1. Models
Let I be the image arrived at decoder. For pixel at site s, I
s ∈ I
,
the average intensity, µ
s, of a local window Gs centered at I
s
is used to calculate the distances between µ
s and the nearest bit
0 and bit 1 quantizers, denoted as d0
s and d1
s respectively, where
d0
s + d1
s = ∆
2 , ∆ is the quantization step for q(.) in QIM. We refer
either d0
s ∀s or d1
s ∀s as a d − map.
In conventional QIM decoder, the bit embedded at site s is decoded
by comparing d0
s with d1
s. The embedded bit is 0 if d0
s ≤ d1
s
and 1 otherwise. Since every bit is decoded from the corresponding
pixel independently, high error decoding rate is inevitable. While
the error rate can be reduced significantly by embedding bits into
local averages instead of original pixels [5], further improvement
is possible by exploiting the contextual information when the bits
in the embedded message (logo here) have some spatial coherency,
e.g. a bit is more likely to be 1 (0) if its neighboring bits are 1(0). In
this paper, we propose a maximum a posteriori probability (MAP)
decoding method to utilize such contextual information.
EXPERIMENTAL RESULTS AND DISCUSSION
3.1. Experimental setup and results
In our experiments, we set the radius for the sliding window r=2,
quantization step ∆=10, β = 1 and σ is set to the half of the
quantization step ∆. ICM was initialized by minimum distance
(MD) decoding. In most cases, ICM converged within 6 iterations.
Notice that MD decoding is equivalent to maximum likelihood
(ML) decoding (i.e. without exploiting any prior information in the
MAP decoding).
Experiments were repeatedly carried out on 25 512×512 images
with 10 different logos. Due to the limited space we only show the
results of embedding a logo (as shown in Fig.3) that is composed
of characters into the well-known image lena (as shown in Fig.2).
It has to be pointed out that the PSNR of the watermarked image
mainly depends on the quantization step ∆ and is about 38.8db at
∆ = 10.
To demonstrate the robustness of the proposed SW-LAQIM
against JPEG compression, we embedded the logo into the image
using both conventional QIM (one bit per pixel as well) and
SW-LAQIM. Both watermarked images were gone through JPEG compression at 50% quality level. Then we employed MD and MAP
decoding to extract the logo from the compressed watermarked
images. Fig.4 shows the decoded logo from the watermarked image
using conventional QIM and Fig.5 shows the decoded logo from watermarked
image using SW-LAQIM. It is obvious that SW-LAQIM
improves the robustness against JPEG compression compared to
conventional QIM and MAP decoding brings substantial further
improvement. The scheme of SW-LAQIM plus MAP decoding
produces clearly identifiable logo.
Fig.6 plots the error-decoding rates of MAP and MD decoding
against the JPEG compression from quality level 90% to 20%
on a SW-LAQIM watermarked lena. The rates were averaged
over 10 logos. At all compression quality levels, MAP decoding
consistently outperforms the MD decoding. The minimum error
rate of MD decoding is above 30% at quality level of 80% and the
error rate of MAP decoding is only around 5% at the same quality
level. Experiments on other images presented similar results.
Fig.7(left) shows the MAP decoded logos on half-sized SWLAQIM
watermarked lena. We also further scale the JPEG 50%
compressed SW-LAQIM watermarked lena to 90% of its original
size, rotate (10 degree counter-clockwise), and cropped to an image
of size 384×384. The decoded logo is shown in Fig.7(right). Although
the logo is more distorted and incomplete, it is still clearly
identifiable.
CONCLUSION
A pixel-based watermarking system has a great advantage of surviving
geometric attacks since the synchronization of the watermark
embedding and decoding is always held. However, a pixel-based
system is usually too fragile for practical application: even the high
quality JPEG compression can remove or disrupt the watermark.
Our proposed system overcomes this problem by utilizing a sliding
window based QIM embedding together with MAP decoding. Experimental
results have demonstrated that MAP decoding performs
substantially better than the Minimum Distance decoding against
JPEG compression and the proposed system presents strong robustness
against JPEG compression and geometric attacks as well.