IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO.
2, FEBRUARY 2014
Adaptive Watermarking and Tree Structure
Based Image Quality Estimation
Sha Wang, Dong Zheng, Jiying Zhao, Member, IEEE, Wa James Tam, and Filippo Speranza
Abstract—Image quality evaluation is very important. In appli-
cations involving signal transmission, the Reduced- or No-Refer-
ence quality metrics are generally more practical than the Full-
Reference metrics. In this study, we propose a quality estimation
method based on a novel semi-fragile and adaptive watermarking
scheme. The proposed scheme uses the embedded watermark to es-
timate the degradation of cover image under different distortions.
The watermarking process is implemented in DWT domain of the
cover image. The correlated DWT coefficients across the DWT sub-
bands are categorized into Set Partitioning in Hierarchical Trees
(SPIHT). Those SPHIT trees are further decomposed into a set of
bitplanes. The watermark is embedded into the selected bitplanes
of the selected DWT coefficients of the selected tree without causing
significant fidelity loss to the cover image. The accuracy of the
quality estimation is made to approach that of Full-Reference met-
rics by referring to an "Ideal Mapping Curve" computed a priori.
The experimental results show that the proposed scheme can esti-
mate image quality in terms of PSNR, wPSNR, JND and SSIM with
high accuracy under JPEG compression, JPEG2000 compression,
Gaussian low-pass filtering and Gaussian noise distortion. The re-
sults also show that the proposed scheme has good computational
efficiency for practical applications.
Index Terms—Bitplane decomposition, complexity analysis,
DWT based watermark embedding, HVS masking, SPIHT tree
structure, watermarking based image quality estimation.
I. INTRODUCTION
T HE evaluation of image quality is very important in
today’s video broadcasting, transmission control, and
e-commerce, because quality is a key determinant of customer
satisfaction and a key indicator of transmission conditions.
Meanwhile, quality is very useful in the evaluation of the
effectiveness or performance of image processing algorithms
or systems. Based on the dependence on a reference image, the
image quality metrics can be divided into three categories: the
Full-Reference, Reduced-Reference and No-Reference quality
metrics. The Full-Reference quality metrics evaluate image
quality by comparing the differences between the distorted
Manuscript received September 27, 2012; revised January 13, 2013; accepted
September 17, 2013. Date of publication November 20, 2013; date of current
version January 15, 2014. The associate editor coordinating the review of this
manuscript and approving it for publication was Dr. Ton Kalker.
S. Wang and J. Zhao are with the School of Electrical Engineering and Com-
puter Science, University of Ottawa, Ottawa, ON K1N 6N5, Canada (e-mail:
shawang@eecs.uottawa.ca; jyzhao@eecs.uottawa.ca).
D. Zheng, W. J. Tam, and F. Speranza are with the Advanced Video Sys-
tems, Communication Research Centre Canada, Ottawa, ON K2H 8S2, Canada
(e-mail: dong.zheng@crc.ca; james.tam@crc.ca; filippo.speranza@crc.ca).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMM.2013.2291658
1520-9210 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
311
image and the original image [1]–[7]. The widely used quality
metrics in this category are the PSNR, wPSNR, Watson JND,
SSIM, etc. The Full-Reference quality metrics provide more
accurate quality evaluation results comparing to the Reduced
or No-Reference quality metrics. However, the Full-Reference
quality metrics become less practical when the original image is
not available. The Reduced-Reference quality metrics evaluate
the quality of a distorted image using partial information of the
original image. In literature, such partial information can be
some features extracted from the original image [8]–[10]. The
Reduced-Reference quality metrics do not require the presence
of the original image for quality evaluation. However, the
partial information of the original image need to be transmitted
to the receiver side either through an ancillary channel or by
embedding into the transmitted image. The sacrifice of band-
width for transmitting the additional information needs to be
considered. The No-Reference quality metrics estimate image
quality without accessing the original image [11]–[16]. In
practical applications, different quality metrics can measure the
image degradation from different angle. For example, PSNR
measures image quality mathematically in terms of MSE; JND
and SSIM intend to measure image quality with more emphasis
on the perceptual experience. Multiple metrics are often used
in practical applications to give a full coverage of the image
quality evaluation. In terms of applicability in signal transmis-
sion, Reduced- or No-Reference metrics are more preferable
over the Full-Reference metrics. To this end, a Reduced- or
No-Reference quality evaluation scheme which can evaluate
image quality in terms of different existing metrics would be
really useful.
The watermarking technique is one of the most promising
methods to develop Reduced or No-Reference quality metrics.
In this case, usually a semi-fragile watermark is embedded in the
original (or cover) image. Both the embedded watermark and
the cover image will undergo the same distortion. The image
quality can be estimated by evaluating the watermark degrada-
tion. The watermarking based quality estimation schemes can be
categorized into the image-feature-dependent schemes and the
image-feature-independent schemes. In the image-feature-de-
pendent schemes, normally, some features extracted from the
original image are embedded into the original image [9], [17],
[18]. At the receiver side, the embedded features are recon-
structed and are used as the reference watermark. The features
extracted from the distorted image are used as the distorted wa-
termark. The watermark degradation is assessed by comparing
the distorted watermark with the reference watermark. Instead
of the original image features, using the reconstructed image
features as the reference watermark may introduce additional in-
accuracy to the quality estimation. Moreover, for these schemes,
312
it is hard to examine which kind of image features are suit-
able for providing accurate quality estimation. The image-fea-
ture-independent schemes simplify the situation [19]–[23]. In
these schemes, the watermark is independent of the image fea-
tures and needs to be known at the receiver side. The water-
mark degradation is evaluated using the distorted watermark
and the original watermark. One challenging task to develop the
image-feature-independent watermarking based quality metric
is to enable the embedded watermark to accurately reflect the
quality changes of cover image under distortions. This requires
the watermark being adaptively embedded in cover images ac-
cording to the characteristics of cover images, so that the em-
bedded watermark degrades in a similar way as the cover images
under distortions. This is a critical part of the whole scheme and
directly affects the accuracy of quality estimation.
In [19], an image-feature-independent watermarking based
quality estimation scheme was proposed which attempts to
achieve the quality estimation accuracy of the Full-Reference
objective metrics. In the scheme, the watermark degradation
is measured using the True Detection Rates (TDR). Then the
image quality is estimated by mapping the calculated TDR to
a quality value using an empirical mapping function which
is experimentally generated and is used as a priori at the
receiver side. An iterative process is used to find the optimal
watermark embedding strength by experimentally testing the
image degradation characteristics so that the quality estimation
error can be minimized. This iterative process provides high
accuracy to the quality estimation. However, it also introduces
relatively high computational complexity which makes it less
suitable for certain applications. Meanwhile, in the scheme,
the human perception characteristics have not been taken into
consideration during the watermark embedding process and
the quality of the watermarked image is about 40 dB in PSNR
on average. Thus, the goal of our research in this paper is
to keep the accuracy of quality estimation achieved in [19]
while improving the computational efficiency and reducing the
image quality degradation caused by the watermark embedding
process. Here, the term “accuracy” evaluates the correlation of
the estimated quality and the quality calculated using the ex-
isting objective Full-Reference quality metrics, such as PSNR.
The closer the estimated quality to the calculated quality, the
more accurate the quality estimation, and vice versa.
In this paper, we propose a new approach that well meets
our research goal addressed above. In the proposed scheme, the
adaptive watermark embedding strength is estimated by ana-
lyzing the quality degradation characteristics of the cover image
and no iterative adjustment loops are used, which significantly
improves the computational efficiency and theoretically makes
the proposed scheme applicable to real-time video quality es-
timation. Moreover, the strategies including the HVS masking
are used to guide the watermark embedding process. With the
proposed scheme, the quality of the watermarked images refer-
ring to the original images is about 48 dB in PSNR on average,
which is an 8-dB improvement over the scheme in [19].
The proposed scheme is based on adaptive watermarking and
tree structure in the DWT domain. Recently, the Set Partitioning
in Hierarchical Trees (SPIHT) has become one of the most pop-
ular image and video coding method because of its efficiency
which is accomplished by exploiting the inherent similarities
IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014
across the subbands in the wavelet decomposed image. The
DWT and SPIHT together provide a good summarization of
local region characteristics of an image which is important for
adaptive watermark embedding. In this proposed scheme, all the
correlated DWT coefficients across the subbands are grouped
together using the SPIHT tree structure. The DWT decomposed
image is further decomposed into a set of bitplane images. In this
case, each DWT coefficient is decomposed into a sequence of
binary bits. The binary watermark bits are embedded into the se-
lected bitplanes of the selected DWT coefficients of the selected
trees. The HVS masking is used to guide the bitplane selection.
As found in the experiments, the higher frequency DWT sub-
bands and less significant bitplanes are more sensitive to dis-
tortions, and vice versa [19]. Therefore, the robustness of the
watermark is controlled by two factors:
(a) The percentages of the watermark bits embedded into the
three DWT levels, respectively, and,
(b) The selection of bitplanes for watermark embedding.
Thus, for different selected trees, the watermark embedding
strengths are different.
The proposed scheme is tested in terms of PSNR, wPSNR,
Watson JND and SSIM, and under JPEG compression,
JPEG2000 compression, Gaussian low-pass filtering and
Gaussian noise addition. The results show the effectiveness of
the proposed scheme.
The rest of the paper is organized as follows. Section II,
Section III and Section IV present the proposed watermark
embedding scheme in detail. Section V describes the watermark
extraction and quality evaluation scheme. Section VI details
the generation of the “Ideal Mapping Curve”. Section VII
shows the experimental results and accuracy evaluations of
the proposed scheme. Section VIII lists the advantages of the
proposed scheme. Section IX concludes the paper and discusses
future work.
II. THE PROPOSED WATERMARK EMBEDDING SCHEME
The proposed watermark embedding scheme is shown in
Fig. 1, which includes the watermark pre-processing, the image
pre-analysis, and the watermark embedding. The watermark
embedding process consists of the following three steps:
(i) Apply 3-level DWT to the original image to obtain the
DWT decomposed image. The 3-level DWT decomposed
subbands are denoted as shown in Fig. 2(a). These denota-
tions for the 10 DWT decomposed subbands will be used
throughout the paper.
(ii) Embed the watermark with adaptive embedding strength
using the tree structure based watermark embedder which
will be presented in Section III in detail. The output of the
watermark embedder is the watermarked DWT image.
(iii) Apply 3-level inverse DWT to the watermarked DWT
image to obtain the watermarked image.
The watermark pre-processing procedure is shown in the
grey-dashed block with number 1 on its upper right corner. In
this procedure, a two dimensional original watermark is orga-
nized column by column into a one dimensional sequence. We
denote the length of the original watermark sequence as len. To
increase the probability of the correct watermark bit extraction
at the receiver side, every bit in the original watermark sequence
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION
Fig. 1. The proposed watermark embedding process.
Fig. 2. Illustration of the 3-level DWT decomposed subbands (a) and the for-
mation of tree structure (b).
is repeated a few times to get a redundant watermark sequence
for watermark embedding. A parameter, , is in-
troduced to express the watermark redundancy. In the proposed
scheme, we set and the original watermark
sequence is repeated times to get the redun-
dant watermark sequence with bits long.
The watermark bit extraction will be mathematically described
in Section V.
The image pre-analysis procedure is shown in the
grey-dashed block with number 2 on its upper right corner. In
this procedure, the texture characteristics and the perceptual
masking effects of the cover image will be analyzed respec-
tively in the spatial and DWT domains. Based on these image
analysis, the adaptive watermark embedding strength is esti-
mated. The details will be presented in Section IV.
The position separation key is used to locate the positions for
watermark embedding.
III. TREE STRUCTURE BASED WATERMARK EMBEDDER
The tree structure based watermark embedder is designed to
embed the binary watermark bits into the selected bitplanes of
the selected DWT coefficients of the selected trees. In the pro-
posed scheme, the tree structure based watermark embedder has
three functions: forming the tree structure, selecting the trees
and the DWT coefficients for the watermark embedding, and
313
embedding the binary watermark bits into the selected bitplanes
of the selected coefficients. All these functions will be presented
in the following three subsections, respectively.
A. The Formation of the Tree Structure
The tree structure is formed by categorizing the DWT coef-
ficients with inherent similarities across all the DWT subbands.
The correlated coefficients build up the parent-descendants re-
lationship and form a tree [24], [25]. An example of the tree for-
mation is shown in Fig. 2(b). In the subband, 1/4 DWT co-
efficients with both odd horizontal and odd vertical coordinates
have no descendants. Only the coefficients with at least one even
coordinate have descendants and we define these coefficients
as the roots of the trees. For the DWT coefficients with odd
horizontal and even vertical coordinates, all their descendants
are located in the detail subbands, where .
For the coefficients with even horizontal and even vertical co-
ordinates, all their descendants are in the detail subbands.
Correspondingly, all the descendants for the coefficients with
even horizontal and odd vertical coordinates are in the
subbands. Each root has descendants in the cor-
responding detail subbands on level , where is the total
number of DWT decomposition levels and in this paper.
Therefore, 3072 trees can be formed for a 512 512 image.
The number of the selected trees can be evenly distributed in
the three orientations, , and , or more in some
orientation(s) and less in the other(s). With the formation of
the tree structure, the locations for the watermark embedding
can be more flexibly chosen. Moreover, the watermark bits can
be embedded into the selected trees with different embedding
strengths.
B. The Selection of Trees and DWT Coefficients
For the applications of the watermarking based quality esti-
mation, it is desirable to embed watermark throughout the cover
image so that, even the watermarked image is locally tampered,
the extracted watermark can still reflect the quality degradation
of the cover image. According to the length of the watermark
sequence, the trees for watermark embedding are chosen using
the position separation key.
314 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014

In the proposed scheme, to keep the embedded watermark

invisible and limit the image quality degradation caused by the
watermark embedding, the watermark bits are not embedded
into the subband of the DWT decomposed image and the
watermark bits are not embedded into the bitplanes higher than
5, where the least significant bitplane is bitplane 1. In this paper,
we denote the watermark bit assignment as ,
where , and are the number of watermark bits to be
embedded in the DWT level 1, 2 and 3 in every selected tree.
For watermark embedding, the redundant watermark sequence
is divided into segments as depicted in (1).

Fig. 3. The tree selection strategy.

(1)
where is the length of the watermark sequence. If we denote
the rows and columns of the DWT image as and , there will
be a total of trees out of trees
selected for the watermarking process. Here, we define another
parameter to control the percentage
of trees selected from the three orientations. With , we can
embed the watermark into the three orientations evenly or by
focusing more on one or two orientations.
To minimize the quality degradation of the cover image
caused by the watermark embedding, two strategies are used
for the tree selection:
(a) The trees selected from the three DWT orientations are
non-overlapping in position.
For example, in Fig. 3(a), we use the square blocks in
, and to indicate the tree positions avail- Fig. 4. The tree selection from the three DWT orientations. (a) Graphical il-
able in the three orientations for selection. The trees , lustration. (b) Experimental illustration.
and shown in Fig. 3(a) are all carrying the spatial
information of block 0 in . Thus, if all the trees A,
B and C are embedded with watermark bits, the quality selected trees segment by segment following the order of
of the top left corner of the image will degrade severely. the selected tree position.
Therefore, if tree is chosen for the watermark embed- After the tree selection, we start to embed the watermark bit
ding, trees B and C will not be selected. In this case, segments into the selected trees referring to . On the th
for an image with trees, there DWT level, the watermark bits are embedded into the DWT
are non-over- coefficients one by one until the number of is reached,
lapping tree positions available for the tree selection. In where .
Fig. 3(b), the ‘X’ marked positions represent the trees se- An experimental example of the tree and DWT coefficients
lected from the three DWT orientations for the watermark selection is shown in Fig. 4(b). All the selected DWT coeffi-
embedding. The ‘ ’ marked positions represent the sep- cients are marked as dark points. The darker points means that
aration between any two selected tree positions. In our the lower significant bitplanes are selected, and vice versa. The
implementation, we use uniform separation, defined as bitplane selection will be presented in Section IV-B. In Fig. 4(b),
. will be output as the the position separation key is 3.
position separation key and will be transmitted to the re-
ceiver side. Here, we set .
C. The Watermark Embedding
(b) The trees are selected throughout the DWT decomposed
image. The binary watermark bits are embedded into the selected
The selected trees are further distributed into the bitplanes of the selected DWT coefficients. Here, we denote the
three orientations, as described by . In the proposed watermark bit as , the DWT coefficient bit on the selected
scheme, we set , which means that bitplane as and the watermarked DWT coefficient bit as .
the trees are evenly selected from the three orientations Then, the watermark bit will be embedded using the following
as illustrated in Fig. 4(a). In this figure, the selected trees equation:
are numerically ordered as . As mentioned
previously, the watermark sequence is divided into if
(2)
segments. Thus, the watermark will be embedded into the if
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION 315

IV. THE ESTIMATION OF ADAPTIVE WATERMARK

EMBEDDING STRENGTH

As mentioned in Section I, the watermark embedding strength

is controlled by the empirical watermark bit assignment and the
bitplane selection, which will be presented in the following two
subsections, respectively.

A. The Empirical Watermark Bit Assignment

The watermark bit assignment is denoted as
. Based on the analysis on image content com-
plexity, we empirically assign watermark bits to the 3 DWT
levels in the selected tree using the following steps:
(a) Analyze the content complexity of the cover image and
calculate a complexity index.
(b) Categorize the test images into different groups according
to their complexity indices.
(c) Assign watermark bits to the 3 DWT levels of the cover
image.
The watermark bit assignment will be the same for all the
selected trees in one image and may be different for different
images.
1) The Analysis of Image Content Complexity: The quad-tree
decomposition based complexity analysis is used in the pro-
posed scheme for a better match with the DWT [26]. For gray
scale images, the intensity difference, , is used to verify
Fig. 5. Illustration of the image content complexity analysis. (a)&(c)&(e) Orig-
whether a further quad-tree decomposition is needed. Here, we inal images. (b)&(d)&(f) Quad-tree decomposed images. (b):
define an intensity difference threshold as . If , . (d): . (f): .
the image or current block will be decomposed into 4 sub-blocks
until or the size of the sub-block reaches 1. Each
quad-tree decomposition is recorded as a decomposition node. Three experimental examples are shown in Fig. 5. These
The depth of the decomposition is denoted as the level of de- quad-tree decomposed images are achieved using the threshold
composition. The content complexity of the cover image is as- , where the maximum intensity value of the
sessed using the following equation [26]: cover image is not bigger than 1. In this case, the brighter the
quad-tree decomposed image, the more complex the cover
(3) image. The complexity indices are listed with the quad-tree
decomposed images.
2) The Watermark Bits Assignment: With the complexity in-
where is the current quad-tree decomposition level; is the
dices, the watermark bits are empirically assigned to the images
highest decomposition level; is the number of quad-tree de-
using the following steps:
composition nodes on level . Then, the calculated complexity
(a) The complexity indices are divided into 6 groups using
values of all the images in our image library are normalized. In
(4). One integer index is associated with one group.
the proposed scheme, the normalized complexity value is used
as the complexity index which locates in [0, 1].
The image content complexity analysis evaluates how much
details that an image contains. A higher complexity value in-
dicates that the image is more complex and the image contains (4)
more detail information. Comparing to a less complex image,
the quality of a more complex image degrades faster against the
same distortion [19]. For this case, to reflect the quality degra-
dation of the cover image, we need to embed more watermark where is the group index; is the complexity
bits into the lower DWT levels of a more complex image. For index. , , , , and are the empirical grouping
a less complex image, we consider to embed more watermark thresholds. These thresholds may be different for different
bits into the higher DWT levels. distortions.
316 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014

(b) With the group indices, the watermark bits are assigned to where
the images using (5).
if
o.w.
if
if
(5) if
for HL blocks
for HH blocks
for LH blocks
and
(b) Background luminance: Intensity variations are less vis-
ible over the brighter and darker areas. The luminance
when
masking is denoted as .
when (6)
when

Recall that, in (1), trees are if

(8)
selected for the watermark embedding, where is o.w.
the length of the redundant watermark sequence. Thus,
(c) Spatial masking or edge proximity: The human eyes are
for images with different complexities, the number of se- more sensitive to noise addition near edges or contours of
lected trees, , and the position separation key, , images. This factor, , is evaluated using the empiri-
may be different. cally scaled local energy of the DWT coefficients in all
detail subbands.

B. The HVS and Bitplane Based Perceptual Masking

The other factor that controls the watermark embedding (9)

strength is the bitplane selection for the watermark embedding. where is a weighting parameter and the suggested value
By decomposing the DWT image into a set of binary bitplanes, for is presented in the following equation [27].
each DWT coefficient is decomposed into a binary bit sequence. if
The less significant bits are more sensitive to distortions than (10)
o.w.
the more significant bits. If we embed the whole watermark
in the least significant bitplane of the DWT image, we obtain (d) Texture sensitivity: Intensity variations in highly textured
the most fragile watermark and the least quality degradation areas are less visible than those in the flat-field areas of
to the cover image, and vice versa. In the proposed scheme, images. This masking factor, , is estimated using the
the bitplane selection is guided by the analysis of the HVS local variance of the corresponding DWT coefficients in
masking effects on the DWT decomposed image. The analysis the LL subband.
of the HVS masking effects helps to increase the capacity for
watermarking without causing more quality degradation to the
cover image.
(11)
The bitplane selection includes two steps: calculating the
where and .
HVS masks and mapping the calculated HVS masks to bit-
The HVS mask is achieved by polling the four factors listed
plane indices. The achieved bitplane indices decide which
above and is computed using (12).
bitplanes of the selected DWT coefficients are used to embed
the watermark bits. The HVS masking calculation and the (12)
HVS-to-bitplane mapping are presented in the following two
subsections. where denotes the HVS mask; is a scaling parameter.
1) The HVS Masking: The HVS masking presented in [27], As analyzed in [27], the suggested value for is 1/2, which
[28] is used in the proposed scheme. Accordingly, four factors implies that intensity variations having values lower than half
greatly affect the behavior of the HVS: of are assumed invisible. The suggested
(a) Band sensitivity or frequency masking: Intensity varia- value for and is 0.2 [27]. In (7) through (12), is
tions are less visible in high resolution subbands and are the luminance value of the pixel, , in one detail subband
also less visible in the diagonally decomposed blocks, on level ; indicates the current DWT
. This factor is expressed using (7). level in the HVS masking calculation and is the max-
imum level applied in the DWT decomposition in this paper;
(7) is the luminance value of pixel
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION 317

Fig. 6. Illustration of the generated HVS masks. (a) The original image Bar-
bara. (b) The nine HVS masks generated for the nine DWT detail blocks.

in the subband which corresponds

to pixel in a detail subband on level .
As proposed in Section III-B, the binary watermark bits are
not embedded in the subband so that the invisibility of the
embedded watermark can be further improved. Therefore, the Fig. 7. Illustration of the thresholds selection. (a)&(d): Original image Bar-
HVS masks are only calculated for the nine detail DWT sub- bara and Treefrog. (b)&(e): Normalized HVS mask of DWT subband HL1 of
(a)&(d). (c)&(f): Histogram of the mask shown in (b)&(e).
bands, with one mask for each. The generated HVS masks for
image Barbara is shown in Fig. 6(b).
2) Mapping From HVS Mask to Bitplane Indices: To
Therefore, each DWT coefficient in the selected trees has its
use the HVS mask in the watermark embedding process, a
own . At the receiver side, the strategy presented
mapping relationship from the coefficients of the HVS mask
above will also be applied on the distorted image to locate the
to the bitplane indices is experimentally defined using the
bitplanes for the watermark extraction. To increase the proba-
multiple-band-thresholding method. Considering that different
bility of correct extraction of the watermark bits, all the calcu-
HVS masks have different distributions, the thresholds used in
lated values on the DWT level in each tree are
the mapping procedure should be able to change with the shape
averaged. In other words, in each selected tree, for all the se-
of the distribution of the HVS mask. To further limit the quality
lected coefficients, the watermark bits will be embedded on the
degradation caused by the watermarking process, only the
same bitplane. Thus, the bitplane indices, , are updated using
bitplanes from 1 to 5 are used for the watermark embedding.
the following equation:
The thresholds calculation and the mapping procedure are
listed as follows:
(a) The HVS mask is first normalized using its maximum co-
efficient so that the mapping relationship can be consistent
with different HVS masks.
(b) The coefficients of the HVS masks are sorted in an as-
cending order and the sorted sequence is denoted as Sort-
where is the averaged bitplane of the DWT
mask.
coefficients located in a specific tree on level and orientation
(c) The thresholds are calculated using (13).
.

(13)

where , is the total number of the

coefficients of one HVS mask, the denominator indicates
that 5 bitplanes are used for the watermark embedding.
Two examples of the thresholds selection are shown in
(d) The mapping is done using (14).
Fig. 7. Fig. 7(b) and (e) are the HVS masks respectively
calculated on the DWT subband HL1 of image Barbara and
image Treefrog. Thus, the size of these two HVS masks
are 1/4 of the image size. Then the HVS-to-bitplane map-
ping thresholds are calculated using (13). Fig. 7(c) and (f)
graphically illustrate the thresholds selection. The com-
(14) puted thresholds for Fig. 7(c) are
where are the coordinates of a selected DWT coeffi- . The thresholds for Fig. 7(f)
cient. means the bitplane index achieved for are .
the pixel located at on DWT level with orientation With the calculated thresholds, the coefficients of the HVS
. is the value of the HVS mask coefficient. mask which correspond to the selected DWT coefficients are
318 IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014

Fig. 8. Watermark extraction and quality estimation.

mapped to the bitplane indices. In this way, the bitplanes of the

selected coefficients for the watermark embedding are located.

V. THE PROPOSED WATERMARK EXTRACTION

AND QUALITY ESTIMATION SCHEME

The proposed watermark extraction and quality estimation

scheme is shown in Fig. 8, where the position separation key
is used to locate the watermarked DWT coefficients.
The watermark bit assignment is retrieved using the image
group index transmitted from the sender side. The bitplane in-
dices for watermark extraction are obtained by calculating the
HVS masks of the distorted watermarked image. As presented in
Section IV-B-2, in one tree, the bitplane indices for all the DWT
coefficients on each DWT level are averaged. This strategy ef-
fectively reduces the watermark extraction error caused by the
bitplane selection in the watermark extraction scheme.
Recall that . The extracted redundant wa-
termark sequence is used to recover the three distorted water-
marks. Then, the three distorted watermarks are compared bit
by bit and the watermark is extracted using (15).

(15) Fig. 9. The “ideal mapping curve” generator.

where is the extracted watermark bit with coordinates

; is the number of extracted 1 s and is the number
of extracted 0 s.
Then, the extracted watermark is compared with the original
watermark bit by bit and the True Detection Rates (TDR) is
calculated using (16).
(16)
As described in (17), the image quality is estimated by map-
ping the calculated TDR to a quality value by referring to a map-
ping function.
(17)
where is the estimated quality; is the mapping function,
which is the “Ideal Mapping Curve” in the proposed scheme.
The generation of the “Ideal Mapping Curve” will be presented
in the next section.
When the calculated TDR is mapped onto the “Ideal Map-
ping Curve”, it could possibly lie between two neighboring TDR
values on the curve. In this case, linear interpolation is used to
estimate the image quality.
VI. THE GENERATION OF “IDEAL MAPPING CURVE”
The “Ideal Mapping Curve” experimentally defines the re-
lationship between the degradation of the watermark and the
quality of the distorted image. The “Ideal Mapping Curve” is
obtained from a set of TDR-Quality curves and is used to map
the calculated TDR to a possible quality value which can be
in terms of the existing Full-Reference quality metric, such as
PSNR, wPSNR, Watson JND or SSIM.
For the image quality estimation in terms of one Full-Ref-
erence quality metric under one type of distortion, one “Ideal
Mapping Curve” needs to be generated. As shown in Fig. 9, the
“Ideal Mapping Curve” is generated by locally averaging the
TDR-Quality curves obtained by testing a number of images.
An example of “Ideal Mapping Curves” generated respec-
tively using 10, 20, 50 and 100 different natural images in
terms of JND under Gaussian low-pass filtering are shown in
Fig. 10. From this figure, we can see that the four generated
“Ideal Mapping Curves” are nearly identical. As mentioned
previously, the adaptive watermark embedding strength is
assigned by analyzing the degradation characteristics of the
cover image. With the adaptive watermark embedding strength,
we expect the TDR-Quality curves tested using a number of
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION
Fig. 10. “ideal mapping curves” generated using different number of test
images.
images as convergent as possible. The more convergent the
tested TDR-Quality curves, the more accurate the quality es-
timation. The similar convergence of the TDR-Quality curves
obtained respectively using 10, 20, 50 and 100 images results
in the four nearly-identical “Ideal Mapping Curves” as shown
in Fig. 10, which also indicates the flexibility of choosing the
number of images for the “Ideal Mapping Curve” generation.
With the proposed scheme, we suggest to use at least 10 natural
images for generating the “Ideal Mapping Curves” to make
it statistically meaningful. In our implementation, we use 50
natural images selected from our image library to generate the
“Ideal Mapping Curves”. The scope of our image library is
presented in Section VII-A.
Two examples of the “Ideal Mapping Curve” gener-
ation under Gaussian low-pass filtering are shown in
Fig. 11. The gray dots are the TDR-Quality curves cal-
culated using the 50 test images. The 50 TDR-Quality
curves are computed when the filter standard deviation
.
When , the PSNR of the low-pass filtered image
is about 25 dB on average. The “Ideal Mapping Curve” is
denoted using black dot dashed curve and is calculated by
locally averaging the 50 gray dot curves. The local averaging
consists of the following steps:
(a) Low-pass filter the test images respectively using the
filter standard deviations listed above. Calculate the
quality values and TDR values under each . Find
the possible quality range, which is [0, 25] JND for
Fig. 11(a) and [20, 49] dB for Fig. 11(b).
(b) List all the desired target quality values in the quality
range found in step (a), which is [0.02, 0.1:0.1:1, 2:1.5:14,
16:2:25] JND for Fig. 11(a) and [49:-0.5:48, 47,46:-1:35
34:-2:20] dB for Fig. 11(b).
(c) Compare each calculated quality with all the target quality
values. By finding the closest target quality, all the calcu-
lated quality values and their corresponding TDR values
are categorized into different groups.
(d) Average the TDR values in each group.
(e) Generate the “Ideal Mapping Curve” using the target
quality values and the averaged TDR values.
319
Fig. 11. Examples of the generation of “ideal mapping curves” under Gaussian
low-pass filtering.
The same averaging rule is used for the “Ideal Mapping
Curve” generation under other distortions. Once the “Ideal
Mapping Curve” is generated, it is stored at the receiver side
and no transmission is needed.
VII. EXPERIMENTAL RESULTS AND ACCURACY EVALUATION
The data need to be transmitted to the receiver side are:
(a) The original binary watermark which is 48 48 pixels
(2304 bits). It is unnecessary to transmit the watermark
if both the sender and receiver sides use the same water-
mark.
(b) The position separation key.
(c) The empirical group index for the watermark bit assign-
ment.
A. The Original Images and the Original Watermark
There are 150 gray images in our image library including all
the reference images from Cornell A57 image database, IVC
image database, TID2008 image database, and CSIQ image
database. Besides these image databases, our image library also
includes computer generated images and more natural images.
All of these images are 512 512 in size and they contain
different textures, such as portraits, plants, animals, sceneries,
buildings, crowd, animation and computer generated patterns.
The first 50 images in the image library are used to generate the
“Ideal Mapping Curves” and other empirical parameters. To
avoid losing generality, the 51st to the 150th images are used
to evaluate the performance of the proposed scheme.
The original binary watermark used in the experiments is 48
48 in size and is randomly generated. The watermark can be
changed to any size or pattern.
Fig. 12 shows the watermarked images of
Fig. 5(a), (c) and (e), respectively. The watermark bit
assignment, , associated with the images are used for
watermark embedding under Gaussian low-pass filtering. The
quality of the three example watermarked images shown in
Fig. 12 are more than 47.5 dB in PSNR, which indicates that
the watermark embedding process does not have significant
impact on the image quality.
B. Summary of the Experiments
The proposed scheme is evaluated under JPEG compres-
sion, JPEG2000 compression, Gaussian low-pass filtering and
Gaussian noise addition. Under each distortion, a watermarked
image is distorted using a variety of distortion strengths and
320
Fig. 12. Illustration of the effects of watermark embedding.
(a) ; Watermark bits assignment:
. (b) ; Watermark bits assign-
ment: ; . (c)
Watermark bits assignment: ; .
the quality of the distorted images is estimated in terms of
PSNR, wPSNR, Watson JND [29] and SSIM [2], respectively.
Therefore, totally 16 sets of experiments are conducted. In
each set of the experiments, 100 images are tested. To assess
the performance of the proposed scheme, the quality of the
distorted image comparing to the original image is calculated
as well. With both the estimated quality and calculated quality,
the quality estimation accuracy is evaluated using MAE (Mean
Absolute Error), Pearson correlation coefficient, and RMSE
(Root Mean Squared Error) respectively. In the experiments,
the image quality degradation caused by both the watermark
embedding and the distortion is considered in the performance
evaluation.
C. Choosing the Range for the Distortion Strength
The watermarked images are distorted using a set of dis-
tortion strengths varying in a selected range, , (f) Set up the empirical thresholds by examining the catego-
where is the distortion strength; and respectively are
the lowest and the highest distortion strength used in the ex-
periments. We choose the lowest distortion strength available
in Matlab as . For example, under JPEG compression, the
quality factor controls the distortion strength and . Gaussian noise distortion with a fixed noise standard deviation,
is so set that the average quality of the watermarked images dis-
torted using , , is in the range of [24.50, 25.50] dB
in PSNR. , where the
is the watermarked image and is the original image. On
the other hand, when the quality of the distorted image is below
30 dB in PSNR, the image quality is poor. The lower the image
quality, the more meaningless to focus on the quality estimation
accuracy. Thus, we choose another high boundary for :
if in PSNR. The 150 images in
the image library are used to choose , and . The range
of the distortion strengths for the four types of distortions used
in the experiments are listed in the following table:
IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014
TABLE I
RANGE OF DISTORTION STRENGTHS AND CORRESPONDING
WATERMARKED IMAGE QUALITIES IN PSNR
In Table I, , and respectively stand for
, and . For JPEG, the distortion
strength is the compression strength and is denoted using
“Quality Factor” in Matlab. For JPEG2000, the distortion
strength is the compression rate. For Gaussian low-pass fil-
tering (denoted as LPF in the table), the distortion strength is
controlled by the standard deviation of the low-pass filter. For
Gaussian additive noise, the distortion strength is the standard
deviation of the Gaussian noise.
The accuracy of the quality estimation under both and
; will be evaluated and shown in Section VII-E. Note that
;
is an averaged quality value. It is normal if some quality
points smaller than appear in the accuracy evaluation
figures.
D. The Empirical Grouping Thresholds
For each set of experiments, the grouping thresholds used in
(4) are empirically set by testing the first 50 images in the image
library using the following steps:
(a) Based on the research done in [19], empirically set up the
watermark bit assignment, , as listed in (5).
(b) With each , test the quality estimation error using the
50 images.
(c) For each test image, find the minimum quality estimation
error and the corresponding optimal .
(d) Calculate the content complexity values for the 50 test
images using the local intensity difference threshold 0.17,
where the maximum intensity value for the images is 1.
(e) Referring to the optimal assigned to the 50 test im-
ages, categorize the calculated content complexity values
into 6 groups.
rized content complexity values.
For different set of experiments, the empirical grouping
thresholds may be different. The thresholds used in the experi-
ments are listed in Table II. As found in the experiments, under
the quality of different distorted images varies slightly no
matter how complex the images are. Thus, under Gaussian
noise distortion, we use the grouping thresholds the same as
those used under JPEG compression for convenience. These
thresholds are stored at the receiver side once they are tested.
E. Experimental Results Under Different Distortions
For the four types of distortions, the distortion strengths used
in the experiments are listed in the following:
• JPEG: ;
• JPEG2000:
.
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION 321

Fig. 13. The generated “ideal mapping curves” in terms of different full-reference quality metrics under different distortions.

TABLE II
THE EMPIRICAL THRESHOLDS USED FOR THE 16 SETS OF EXPERIMENTS
• Low-pass filtering:
. the accuracy of the quality estimation when
• Gaussian noise distortion:
. the calculated quality is smaller than 8 JND is evaluated as well.
For the 16 sets of experiments, the 16 generated “Ideal
Mapping Curves” are shown in Fig. 13(a)–(p). For example,
Fig. 13(a), (e), (i) and (m) show the “Ideal Mapping Curves”
for the quality estimation in terms of PSNR. With increasing
distortion strength, the calculated PSNR value and the calcu-
lated TDR value decrease monotonically.
Fig. 13(c), (g), (k) and (o) are the “Ideal Mapping Curves”
used to estimate image quality in terms of JND. With increasing
distortion strength, the calculated TDR decreases and the cal-
culated JND value increases. When TDR equals to 1, the JND
is 0. In Fig. 13(c), is a slope change point and
JND, where is the quality. When
, the JND value increases rapidly with the decreasing TDR
value. When TDR decreases from 0.4 to 0.31, the JND value in-
creases from 8 to 40. Thus, we need to use a TDR value varying
in a 0.09 value range to estimate image quality varying in a 32
JND value range. In this case, a small variation of the TDR value
may result in a relatively significant quality estimation error. So,
we can predict that this rapid increase of JND will greatly reduce
. For
additional observations, the quality estimation accuracy when
Similar analysis can be made to Fig. 13(g), (k), and (o).
Fig. 13(d), (h), (l) and (p) are the “Ideal Mapping Curves”
used to estimate image quality in terms of SSIM. In Fig. 13(d),
the left half of the curve has a very big slope. When the TDR
decreases from 0.42 to 0.31, the calculated SSIM changes from
0.9871 to 0.7. The SSIM values drop rapidly corresponding a
small change of TDR value. It is straightforward that a small
variation of the TDR value may cause relatively bigger estima-
tion error and the divergence of the quality points referring to
the solid matching line may be magnified.
The 16 sets of experimental results of image quality estima-
tion tested with distortion strengths are shown in
322
Fig. 14. The experimental results tested with distortions in terms of different Full-Reference quality metrics under different distortions.
Fig. 14(a)–(p). In the figures, the x-axis is the estimated quality.
The y-axis is the correspondingly calculated quality. The black
points are the quality points achieved under different distortion
strengths by testing the 100 images. The solid line is the quality
matching line with a slope of 1, which indicates that the esti-
mated quality equals to the calculated quality. The discrepancy
from the solid line to the quality points indicate the accuracy of
the quality estimation. The more convergent the quality points
to the solid line, the more accurate the quality estimation and
vice versa. If all the quality points fall on the solid line, the es-
timated quality equals to the calculated quality and the highest
quality estimation accuracy is achieved. For each sub-figure in
Fig. 14, the corresponding quality estimation accuracy in MAE
is listed. Recall that in PSNR. When the
quality of a distorted image is lower than 30 dB in PSNR, the
image quality becomes poor and the quality points become more
IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014
scattering. The poorer the image quality, the more meaning-
less to focus on the quality estimation accuracy. The accuracy
of the image quality estimation evaluated using MAE, Pearson
correlation coefficient and RMSE is summarized in
Table III, Table IV and Table V, respectively. In the tables,
and respectively indicate the range of the distortion strengths
and , which are described in Section VII-C and
represents stronger distortion strength than . LPF stands
for low passing filtering, and Gaussian stands for Gaussian noise
addition.
From Table III through Table V, we can see that:
(a) The proposed scheme has good accuracy to estimate
image quality in terms of PSNR and wPSNR. Under
the four mentioned distortions, with both the distortion
strengths and , the calculated for
the quality estimation in PSNR are respectively lower
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION
TABLE III
SUMMARY OF THE ACCURACY IN MAE
TABLE IV
SUMMARY OF THE ACCURACY IN PEARSON CORRELATION
TABLE V
SUMMARY OF THE ACCURACY IN RMSE
than 0.74 dB and 0.8710 dB; and the calculated
for the quality estimation in wPSNR are respectively
lower than 0.77 dB and 0.9405 dB. The gives
similar results as the does. The calculated Pearson
correlations are all higher than 0.977 which indicates the
high correlations between the estimated quality and the
calculated quality.
(b) To estimate image quality in terms of JND, the proposed
scheme has the highest accuracy under Gaussian low-pass
filtering with the distortion strength in both the range
and . Under other distortions, the proposed
scheme has good quality estimation accuracy with
and acceptable quality estimation accuracy with
. accuracy. Moreover, it can be used to evaluate image
(c) To estimate image quality in terms of SSIM, the proposed
scheme has good quality estimation accuracy under all
the four distortions with and acceptable
accuracy with . posed scheme can be developed to estimate image quality
Under JPEG compression, the performance of the proposed
scheme is compared with that of the scheme in [19]. Our
previous scheme [19] was tested by estimating image quality
in terms of PSNR, wPSNR, JND under JPEG compression and
achieved good accuracy. The quality factors [100:-5:20] are
used in [19] for the image quality estimation. The accuracy
of the image quality estimation of both our previous scheme
and the currently proposed tree structure based scheme under
JPEG compression with quality factors [100:-5:20] are listed
in Table VI for comparisons. From both the calculated MAE
and the calculated Pearson correlations, , we can see that
the two schemes can provide comparable accuracy for image
quality estimation.
323
TABLE VI
COMPARISON BETWEEN THE ACCURACY OF OUR PREVIOUS
SCHEME AND THE PROPOSED SCHEME UNDER JPEG COMPRESSION
In the proposed scheme, four sets of parameters and thresh-
olds have been introduced: the grouping thresholds for calcu-
lating in (4), the watermark bit assignment parameters
in (5), the parameters used to calculate the HVS mask
in (12), and the mapping parameters from HVS masks to
bitplane indices in (14). The first set of parameters in (4) is
image independent and distortion dependent. The rest three sets
of empirical parameters are image and distortion independent,
which means that they are fixed for all images and all types of
distortions. All these parameters and thresholds work together
to balance the robustness of algorithm and the quality degrada-
tion caused by the watermark embedding.
VIII. ADVANTAGES OF THE PROPOSED SCHEME
Based on the performance evaluation in Section VII-E, the
advantages of the proposed adaptive watermarking and tree
structure based quality metric are listed in the following:
(a) Computational efficiency. We have mentioned that our
previous work [19] requires an average of 12 recursive
loops to find the optimal watermark embedding strengths
for the test images. With the proposed scheme, the exten-
sive iterative process is not needed and the computational
efficiency is improved about 17 times over that of our pre-
vious work [19] based on the computation cycle counting
using profiling tools.
(b) The quality loss caused by the watermark embedding is
very small. For 150 images, the PSNR of the watermarked
image is 48.1476 dB on average.
(c) The proposed scheme can assess image quality in terms
of PSNR, wPSNR, Watson JND and SSIM with good
quality degradation under different distortions, such as
JPEG compression, JPEG2000 compression, Gaussian
low-pass filtering and Gaussian noise addition. The pro-
in terms of other Full-Reference quality metrics and or
under other distortions.
(d) It works as a Reduced or No-Reference quality evaluation
scheme which is more desirable in communication sys-
tems. With the proposed scheme, the cover image is un-
necessary to be transmitted to the receiver side for quality
evaluation, which greatly saves the transmission band-
width.
IX. CONCLUSIONS AND FUTURE WORK
In this paper, a new watermarking based quality estimation
scheme is presented. The proposed scheme is designed to es-
timate image quality in terms of the existing Full-Reference
324
quality metrics, such as PSNR, wPSNR, JND and SSIM. Thus,
at the receiver side of a communication system, without the orig-
inal image, the quality of a distorted image can still be assessed.
Based on the tree structure, the binary watermark is em-
bedded into the selected bitplanes of the selected DWT
coefficients with adaptive watermark embedding strength.
The watermark embedding strength is assigned to an image
by pre-analyzing its content complexity in the spatial domain
and the perceptual masking effect of the DWT decomposed
image in the DWT domain. Meanwhile, the watermark is not
embedded in the approximation subband, which reduces loss
in image quality caused by embedding the watermark. The
experimental results show that the proposed scheme works
effectively.
In future work, the proposed scheme will be further devel-
oped to estimate the quality of an image distorted by multiple
distortions. Meanwhile, experiments about image quality esti-
mation in terms of subjective quality scores will be conducted.
Since the proposed scheme has good computational efficiency,
it is feasible to further develop the proposed scheme for video
quality evaluation.
REFERENCES
[1] I. J. Cox, M. L. Miller, J. A. Bloom, J. Fridrich, and T. Kalker, Digital
Watermarking and Steganography. San Mateo, CA, USA: Morgan
Kaufmann, 2008.
[2] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image
quality assessment: From error visibility to structural similarity,” IEEE
Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004.
[3] S. Wolf, M. Pinson, S. Voran, and A. Webster, “Objective quality as-
sessment of digitally transmitted video,” in Proc. IEEE Pacific Rim
Conf. Communications, Computers and Signal Processing, 1991, vol.
2, pp. 477–482.
[4] A. Watson, Q. Hu, and J. McGowan, “Digital video quality metric
based on human vision,” J. Electron. Imag., vol. 10, pp. 20–29, 2001.
[5] A. Shnayderman, A. Gusev, and A. Eskicioglu, “An SVD-based
grayscale image quality measure for local and global assessment,”
IEEE Trans. Image Process., vol. 15, no. 2, pp. 422–429, 2006.
[6] G. Zhai, W. Zhang, X. Yang, S. Yao, and Y. Xu, “GES: A new image
quality assessment metric based on energy features in Gabor transform
domain,” in Proc. IEEE Int. Symp. Circuits and Systems, 2006, pp.
1715–1718.
[7] M. Carnec, P. Callet, and D. Barba, “Full reference and reduced ref-
erence metrics for image quality assessment,” in Proc. 7th Int. Symp.
Signal Processing and Its Applications, 2003, vol. 1, pp. 477–480.
[8] Q. Li and Z. Wang, “Reduced-reference image quality assessment
using divisive normalization-based image representation,” IEEE J.
Select. Topics Signal Process., vol. 3, no. 2, pp. 201–211, 2009.
[9] Z. Wang, G. Wu, H. R. Sheikh, E. P. Simoncelli, E. H. Yang, and A. C.
Bovik, “Quality-aware images,” IEEE Trans. Image Process., vol. 15,
no. 6, pp. 1680–1689, 2006.
[10] T. Kusuma, H. Zepernick, and M. Caldera, “On the development of a
reduced-reference perceptual image quality metric,” in Proc. Systems
Communications, 2005, pp. 178–184.
[11] P. Lecallet, C. Viard-Gaudin, S. Pechard, and E. Caillault, “No refer-
ence and reduced reference video quality metrics for end to end QoS
monitoring,” IEICE Trans. Commun., vol. E85-B, no. 2, pp. 289–296,
2006.
[12] D. Turaga, Y. Chen, and J. Caviedes, “No reference PSNR estimation
for compressed pictures,” Signal Process.: Image Commun., vol. 19,
no. 2, pp. 173–184, 2004.
[13] P. Marziliano, F. Dufaux, S. Winkler, and T. Ebrahimi, “A no-reference
perceptual blur metric,” in Proc. IEEE Int. Conf. Image Processing,
2002, vol. 3, pp. 57–60.
[14] W. Ci, H. Dong, Z. Wu, and Y. Tan, “Example-based objective quality
estimation for compressed images,” IEEE Multimedia, to be published.
IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 16, NO. 2, FEBRUARY 2014
[15] H. Wu and M. Yuen, “A generalized block edge impairment metric for
video coding,” Signal Process. Lett., vol. 4, no. 11, pp. 317–320, 1997.
[16] F. Battisti, M. Carli, and A. Neri, “No-reference quality metric for color
video communication,” in Proc. Int. Workshop Video Processing and
Quality Metrics, 2012, pp. 1–6.
[17] A. Bhattacharya, S. Palit, N. Chatterjee, and G. Roy, “Blind assessment
of image quality employing fragile watermarking,” in Proc. IEEE 7th
Int. Symp. ISPA, 2011, pp. 431–436.
[18] S. Altous, M. K. Samee, and J. Gotze, “Reduced reference image
quality assessment for jpeg distortion,” in Proc. IEEE 7th Int. Symp.
ELMAR, 2011, pp. 97–100.
[19] S. Wang, D. Zheng, J. Zhao, W. J. Tam, and F. Speranza, “An image
quality evaluation method based on digital watermarking,” IEEE
Trans. Circuits Syst. Video Technol., vol. 17, no. 1, pp. 98–105, 2007.
[20] D. Zheng, J. Zhao, W. Tam, and F. Speranza, “Image quality mea-
surement by using digital watermarking,” in Proc. IEEE Int. Workshop
Haptic, Audio and Visual Environments and Their Applications, 2003,
pp. 65–70.
[21] M. Farias, M. Carli, and S. Mitra, “Objective video quality metric based
on data hiding,” IEEE Trans. Consum. Electron., vol. 51, no. 3, pp.
983–992, 2005.
[22] E. Nezhadarya, Z. Wang, and R. Ward, “Image quality monitoring
using spread spectrum watermarking,” in Proc. IEEE Int. Conf. Image
Processing, 2009, pp. 2233–2236.
[23] A. Mishra, A. Jain, M. Narwaria, and C. Agarwal, “An experimental
study into objective quality assessment of watermarked images,” Int.
J. Image Process., vol. 5, no. 2, pp. 199–219, 2011.
[24] D. Taubman and M. W. Marcellin, JPEG2000: Image Compres-
sion Fundamentals, Standards and Practice. Norwell, MA, USA:
Kluwer, 2002.
[25] A. Said and W. Pearlman, “A new, fast and efficient image codec based
on set partitioning in hierarchical trees,” IEEE Trans. Circuits Syst.
Video Technol., vol. 6, no. 3, pp. 243–250, 1996.
[26] M. Jamzad and F. Yaghmaee, “Achieving higher stability in water-
marking according to image complexity,” Scientia Iranica J., vol. 13,
no. 4, pp. 404–412, 2006.
[27] M. Barni, “Improved wavelet-based watermarking through pixel-wise
masking,” IEEE Trans. Image Process., vol. 10, no. 5, pp. 783–791,
2001.
[28] A. Lewis and G. Knowles, “Image compression using the 2-D wavelet
transform,” IEEE Trans. Image Process., vol. 1, no. 52, pp. 244–250,
1992.
[29] [Online]. Available: http://vision.arc.nasa.gov/dctune/(last visited in
Sep. 2011).
[30] D. Sheskin, Handbook of Parametric and Nonparametric Statistical
Procedures, 2nd ed. London, U.K.: Chapman & Hall CRC, 2000.
Sha Wang received her B.Eng in thermal power
engineering from North China Electric Power Uni-
versity, China in 2002, and received her Master’s
and Ph.D. degree in electrical engineering from
University of Ottawa, Canada, respectively in 2005
and 2013. Her research interests are on image and
video processing.
Dong Zheng received his B.E. degree in electrical
engineering from Wuhan University, China in 1996
and received his M.A.Sc and Ph.D. in electrical engi-
neering from University of Ottawa, Canada respec-
tively in 2003 and 2008. He worked as a senior re-
search engineer in Dilithiumnetwork Inc., California
from 2007 to 2008. He worked as a research scientist
at Communication Research Centre of Canada from
2008 to 2012. Since 2012, he has been working as
a staff software engineer of Magnum Semiconductor
on video processing. His current research interests in-
clude digital video and image coding and quality assessment.
WANG et al.: ADAPTIVE WATERMARKING AND TREE STRUCTURE BASED IMAGE QUALITY ESTIMATION 325

Jiying Zhao (M’00) received the Ph.D. degree
in electrical engineering from North China Elec-
tric Power University, and the Ph.D. degree
in computer engineering from Keio University.
He is a professor with the School of Electrical
Engineering and Computer Science, University of
Ottawa, Ottawa, Canada. His research interests
include image and video processing and multi-
media communications. Dr. Zhao is a member
of the Institute of Electronics, Information and
Communication Engineers (IEICE) and is also a
member of the Professional Engineers Ontario (PEO).
Filippo Speranza received his Ph.D. in Experi-
mental Psychology from the University of Toronto
(Canada) in 1999. From 2001 to 2013, he worked
as Research Scientist with Industry Canada—Com-
munications Research Centre (CRC), where he
investigated stereoscopic imaging and subjective
picture quality assessment techniques and methods.
He is an adjunct professor at the University of
Ottawa.

Wa James Tam received a Ph.D. degree in experi-

mental psychology from York University in Toronto,
Canada in 1990. His graduate thesis was on the
effect of visual fixation on binocular eye movement
latencies. He embarked on his first project, on video
coding based on viewer’s eye movement patterns,
at the Communications Research Center (CRC) in
Ottawa, Canada in 1989. From then on he became
a senior research scientist and a project leader on
3D with the Advanced Video Systems Group at
CRC. After 23 years at CRC, he has moved on to
consultation work on 3D and applied human visual perception. Currently, he is
an adjunct professor at the University of Ottawa.