ArticlePDF Available

Multi-view 3D CG Image Quality Evaluation Including Visible Digital Watermarking Based on Color Information

Authors:

Abstract and Figures

Thus far, we have studied multi-view 3D image quality evaluation, including visible digital watermarking, in the case of considering coded degradation, region, resistance, and number of viewpoints. As a result, the more we process watermarking in the low frequency domain, the more the assessment tends toward independence from image or video patterns. However, we do not consider color information in the case where we generate digital watermarking images. In this paper, first, for the 3D CG images encoded and decoded by H.265/HEVC with 8 Viewpoints Parallax Barrier Method, we process the wavelet transformation, and perform embedding for the arrangement of the high and low frequency domain (LL3, HL3, LH3) in the same frequency level per viewpoint. Next, we evaluate the generated image, analyze the results, and classify for RGB pattern and coded degradation using SVM (Support Vector Machine).
Content may be subject to copyright.
The Eighth International Workshop on Image Media Quality and its Applications, IMQA2016
March 10-11, 2016, Nagoya, Japan
MULTI-VIEW 3D CG IMAGE QUALITY EVALUATION INCLUDING
VISIBLE DIGITAL WATERMARKING BASED ON COLOR INFORMATION
Norifumi Kawabata and Masaru Miyao
†‡Graduate School of Information Science, Nagoya University
E-mail: {norifumi, miyao}@nagoya-u.jp
ABSTRACT
Thus far, we have studied multi-view 3D image quality
evaluation, including visible digital watermarking, in the
case of considering coded degradation, region, resistance,
and number of viewpoints. As a result, the more we
process watermarking in the low frequency domain, the
more the assessment tends toward independence from
image or video patterns. However, we do not consider
color information in the case where we generate digital
watermarking images. In this paper, first, for the 3D CG
images encoded and decoded by H.265/HEVC with 8
Viewpoints Parallax Barrier Method, we process the
wavelet transformation, and perform embedding for the
arrangement of the high and low frequency domain (LL3,
HL3, LH3) in the same frequency level per viewpoint.
Next, we evaluate the generated image, analyze the
results, and classify for RGB pattern and coded
degradation using SVM (Support Vector Machine).
Index Terms― Multi-view 3D Image, H.265/HEVC,
Frequency Domain, RGB Color Space, Double Stimulus
Impairment Scale (DSIS), Support Vector Machine
(SVM)
1. INTRODUCTION
Using QFHDTV (Quad Full HDTV), which has image
resolution four times that of FHDTV (Full HDTV), we
can see high-definition video. Therefore, 3D video can be
of high definition, and this is re-focused as added value.
On the other hand, it is advanced for the research and
development of multi-view 3D image quality assessment
toward the near future. By incorporating high image
quality into smartphones, users can easily see and create
image or video content, and add and embed different
information. Therefore, there is often a problem of
copyright or security.
Previous research has indicated that it is
important for digital watermarking to preserve embedded
watermarking information permanently. It is important
for watermarking information not to change in quality or
disappear from processing data compression [1, 2].
Therefore, research on digital watermarking should
include coded degradation. For visible digital
watermarking, using frequency transformation is
particularly important [37], as is for the research to
consider quality degradation [3].
We perform multi-view 3D image quality
evaluation, including visible digital watermarking, by
changing the coded degradation region by H.265/HEVC,
or the number of viewpoints [8]. Next, we perform multi-
view 3D image quality evaluation, including the
arrangement of visible digital watermarking, by
frequency domain transformation [9]. However, we did
not consider color information in a previous study. From
past studies, it is known that monochrome images have
some resistance; however, color images are vulnerable
for digital watermarking [4]. It has already been reported
that it is difficult for multi-view 3D images to perceive
cases where we embed watermarking for low frequency
components of multi-view 3D images. However, the
relationship between high or low frequency domain and
the RGB component of watermarking images is not clear.
In order to have many viewpoints and parallax in multi-
view 3D, it is not clear how to make the assessors
acceptable by the watermarking method for an image.
Therefore, we need to consider these problems, including
coded image quality and the RGB component of
watermarked images. Past studies have indicated that
digital watermarking relates to multimedia quality, and
there is a problem of how to make the assessors
acceptable by the increase and decrease of multimedia
quality. This problem depends on utilization form or
quality requirement. In general, in multi-view 3D, we
consider that we need that it is clearly for the relationship
between watermarking and media quality.
In this paper, first, we perform wavelet
transformation, and then embed each component of the
watermarked image by RGB decomposition, including
three-level high and low frequency components for the
3D CG image encoded and decoded by H.265/HEVC
with 8 Viewpoints Parallax Barrier Method. Next, we
perform a subjective evaluation experiment for the
generated image and analyze the results. Finally, we
classify the evaluation score for the RGB pattern and
coded image quality using SVM (Support Vector
Machine).
Fig. 1: 3D CG image and watermarking image
2. VISIBLE DIGITAL WATERMARKING AND
SUBJECTIVE EVALUATION EXPERIMENT
2.1. 3D CG image contents used in this study
For this study, we used the 3D CG images shown in Figs.
1 (a)(d) called “Museum” and “Wonder World,”
provided from NICT (National Institute of Information
and Communications Technology) free of charge [10].
On the other hand, we used the watermarking images
shown in Figs. 1 (e)(f) [11]. Figs. 1 (g)–(r) show an
example of generating a watermarking image. We
generated CG content using the procedure shown in Fig.
2. First, we constructed eight-viewpoint CG cameras in
the 3D CG space using an Autodesk Maya plug-in [12].
Then, we processed the camera work and rendering. As a
result, we were able to generate eight viewpoint CG
cameras.
Fig. 2: An example of flowchart and block diagram to generate 3D CG image
Fig. 3: Hierarchical octave division
Next, we encoded and decoded the generated images by
H.265/HEVC, and processed watermarking for the coded
images. Subsequently, we composed eight-viewpoint
images and obtained the evaluation image. We describe
the process for watermarking in subsection 2.2.
For this experiment, we prepared 120 sequences,
including QP (Quantization Parameter) by H.265/HEVC,
the watermarking arrangement of components by RGB
decomposition, and content. We set up the embedding
intensity of the watermarking image to Q=0.05, and the
watermarking arrangement in the frequency domain to
level three.
2.2. Method of embedding based on color information
Figure 2 shows the embedding of a watermarking image
in this experiment. We process watermarking using a
frequency domain transformation (wavelet
transformation) that includes color information. The
generation procedure is as follows:
(1). Decompose the original and watermarking images
into R, G, and B components.
(2). Perform DWT (Discrete Wavelet Transformation)
for the R, G, and B component images generated by
RGB decomposition. For this study, we performed
wavelet transformation using the filter bank 5/3 tap
SSKF (Simple Short Kernel Filter). Figure 3 shows
the ten-bandwidth sub-band region obtained by
three-level octave division. In general, the LL signal
is the low frequency component for the horizontal
and vertical directions, and shows an approximate
image of the image signal in order to focus the
image signal energy on the low component [5]. On
the other hand, sub-band LH, HL, and HH signals
show the horizontal, vertical, and diagonal direction
components, respectively.
(3). For an image by sub-band division, as shown in Fig.
3, we designated each the difference region for R,
G, and B components. Then, embed the
watermarking onto the sub-band image by a linear
combination of the image. After the watermarking
process, sub-band image is obtained as indicated
in Eq. (1).
= +


(1)
Table 1: Main specification of subjective evaluation
experiment
Type of display
Newsight 3D Display 24V type
Display
resolution
1920
×
1080 (Full HD) (pixel)
Display
resolution in the
case of
presentation
1920
×
960 (pixel)
Types of 3D
CG images
“Museum” (M),
“Wonder World” (W)
Types of image
Windows bitmap
Encoding and
decoding
H.265/HEVC
Quantization
Parameter (QP)
 = 0 (), 20,30,40,51
Types of digital
watermarking
images
“Port view” (iheval-04, Pv),
“Flower pot” (iheval-06, Fp)
Digital watermarking
by wavelet transformation
Frequency
domain
component
“LL3”
Embedding
intensity
= 0.05
Embedding
arrangement
“LL,” “HL,” “LH”
3D system
Parallax Barrier Method
The number of
viewpoints
Only 8 viewpoints
Visual range
3H (Height of an image)
Indoor lighting
None (same as dark room)
Assessor’s
position
Within horizontally ±30°
from the center of the screen
Evaluation
experiment
Presentation
time
10
seconds/Contents
Evaluation
method
Double Stimulus
Impairment Scale
Assessor
Repeating 5 times
by an assessor
Here, is the horizontal direction of the image
signal, is the vertical direction of the image signal,
is the coded image, is the watermarking image,
and is the weight coefficient, that is, embedding
intensity of the watermarking image. After
completing the processing, for each sub-band region,
we set up as flat and the weight coefficient of
as one.
(4). Compose the watermarking resulting image of the
R, G, and B components obtained by IDWT
(Inverse Discrete Wavelet Transformation). Finally,
the generated image is obtained.
2.3. Experimental contents
In this study, we performed the three evaluation
experiments described in this subsection.
Fig. 4: Appearance of Fig. 5: Pixels and viewpoints
3D display arrangement of 3D display
Fig. 6: Double Stimulus Impairment Scale
Table 2: EBU method
Score
Quality
5
Imperceptible
4
Perceptible, but not
annoying
3
Slightly Annoying
2
Annoying
1
Very Annoying
Exp. 1: R is in the low frequency domain “LL3.”
Exp. 2: G is in the low frequency domain “LL3.”
Exp. 3: B is in the low frequency domain “LL3.”
As an example, in the case of Exp. 1, G and B are in “HL3”
or “LH3.” We decided not to include multi-components
in the same frequency domain. Exps. 2 and 3 are similar
to Exp. 1. Table 1 lists the experiment specifications. The
experiment environment is based on the ITU-R BT.500-
13 recommendation [14] and reference [1]. In this study,
we conducted our experiment in a dark room in order not
to affect evaluation by luminance and include color
information in the experiment. We used a parallax barrier
display, as shown in Fig. 4 [15]. Figure 5 shows the pixel
arrangement in the parallax barrier display. In the pixel
arrangement, each viewpoint is arranged consistently as
R, G, and B in the direction of the right diagonal. In the
3D display mechanism, the images can be seen in the 3D
display from different fields of view by dividing R, G,
and B through the parallax barrier in front of an LCD
display. When we displayed a 3D image on the 3D
display, we performed real-time parallax mix
automatically using a media player provided by
Newsight Corporation [16]. Furthermore, we defined the
CG camera interval based on the experiment results
reported in [17].
2.4. Experimental method and evaluation
Figure 6 shows the presentation order and image
sequence time for the experiment (DSIS: Double
Stimulus Impairment Scale). First, we displayed
reference image A for 10 seconds, and then mid-gray
image G for 3 seconds. Next, we displayed test condition
image B for 10 seconds. Subsequently, the assessor
evaluated this cycle and inputted the assessment value
into the computer application (VOTE), which required 10
seconds. This cycle was defined as one set, and we
repeated the cycle until the last set. When the assessors
performed the subjective evaluation for longer than 30
minutes, they suffered from accumulated fatigue (such as
visual fatigue) [17—19]. Therefore, when the subjective
evaluation time was longer than 30 minutes, we divided
the assessors’ subjective evaluation time into intervals of
several seconds each. In this experiment, the assessors
inputted the evaluation scores using a computer
application. Table 2 lists the evaluation standard
(European Broadcasting Union: EBU method) [16]. The
assessors assigned evaluation scores according to five
ranks (MOS: Mean Opinion (,)). Here, we
defined  = 4.5, 3.5, and 2.5 as “DL (Detective
Limit),” “AL (Acceptability Limit),” and “EL
(Endurance Limit),” respectively. For this study, each
assessor repeated the evaluation five times. The image
sequences were displayed randomly in each experiment.
We included a rest time of 30 minutes. Thus, we could
reach an appropriate conclusion.
3. EXPERIMENTAL RESULTS
Figures 7–12 show each line graph for MOS (“Pv,” “Fp”)
in Exps. 1, 2, and 3. The vertical line of Figs. 7–12 is
 (,). The horizontal line of Figs. 7–12 is
the embedding pattern of the RGB components (LL3,
HL3, and LH3) and QP. The error bar is extended
vertically from the plot points in Figs. 7–12, which shows
95% confidence interval.
3.1. Result of Exp. 1
3.1.1. “Port view” case
Figure 7 shows that in the RGB case, when  <40 is
satisfied,  and  satisfy “AL;” however, when
 40 is satisfied,  and  decline rapidly,
and when  =40 is satisfied,  and  are
above and below “EL.” On the other hand, in the RBG
case,  is greater than “AL” when QP 40 is
satisfied. Therefore, this result is higher than the case of
RGB.  tends to be the same as for RGB. From this,
we can see the difference in content.
3.1.2. “Flower pot” case
Figure 8 shows that in the case of all patterns for RGB
and RBG,  and  are lower than “Pv.” When
 =40 is satisfied,  and  are lower than
“EL.” Overall, the error range for “Fp” is smaller than for
“Pv.” We can see the difference in content when  =
,20 are satisfied. However, in other QP patterns, we
cannot see this.
3.2. Result of Exp. 2
3.2.1. “Port view” case
Figure 9 shows that in the GRB case, when compared
with Exp. 1,  is the same as for Exp. 1. The
difference between the maximum and minimum for
 is higher than for Exp. 1 by approximately one
when   30 is satisfied. “Pv” mainly includes
the blue color system; therefore, we consider that it is
easy to perceive if we apply a watermark when the low
frequency component is G and not R. In addition, we
consider relating that the background color in “Wonder
World” is nearly black. On the other hand, in the GBR
case, when compared with Exp. 1,  decreases from
“AL” to “EL” when  =40 is satisfied. This tendency
is not seen in the case of GRB. In addition to the low
frequency component becoming G from R, the high
frequency component includes B. We consider that this
is the factor of change.
3.2.2. “Flower pot” case
Figure 10 shows that in the case of GRB and GBR,
   is satisfied most frequently; therefore,
this tends to be the difference when compared with
Exp. 1. Because  and  decrease for both GRB
and GBR, we can consider as one of factors that the low
frequency component changes from R to G.
3.3. Result of Exp. 3
3.3.1. “Port view” case
Figure 11 shows the tendency difference between Exps. 1
and 2. In the BRG case,  > 4.5 (“DL”),  > 4
(“AL”) are both satisfied when  30 is satisfied.
Furthermore, in the BGR case,  and  are
higher (“DL”) than BRG when  30 is satisfied. On
the other hand, the change for both  and  is
larger when  >30 is satisfied, and we can see that the
difference ranges from two to three.
Fig. 7:  in Exp. 1 (Port view) Fig. 8:  in Exp. 1 (Flower pot)
Fig. 9:  in Exp. 2 (Port view) Fig. 10:  in Exp. 2 (Flower pot)
Fig. 11:  in Exp. 3 (Port view) Fig. 12:  in Exp. 3 (Flower pot)
From this, in the case where the low frequency
component includes B, when compared with the others
(R, G), the assessment value tends to be high when coded
degradation does not occur. However, when coded
degradation occurs, it is easy for the assessor to perceive
the coding noise. Therefore, the assessment value tends
to decline rapidly.
3.3.2. “Flower pot” case
Figure 12 shows that in the BRG and BGR cases, 
and  tend to be high, similar to “Pv.” However, the
assessment value is not sufficiently stable to decline
slowly when   30 is satisfied. Compared
with “Pv,” “Fp” consists of colorful components. We
consider this to affect the evaluation results.
3.4. Statistical learning by SVM
Tables 3 and 4 (Exp. 1), 6 and 7 (Exp. 2), and 9 and 10
(Exp. 3) list the SVM results for QP and RGB embedding
by SMO (Sequential Minimal Optimization) [20, 21].
“Precision,” “Recall,” and “F-Measure” values greater
than 0.7 are denoted in boldface font. SVM correctly
classifies percentages for QP and RGB embedding by the
Table 3: SVM of Exp. 1 (“Pv”)
Precision
Recall
F-Measure
Class (QP)
0.35
0.80
0.49
QP_ref
0
0
0
QP_20
0.27
0.30
0.29
QP_30
1
0.60
0.75
QP_40
1
1
1
QP_51
0.52 0.54 0.50
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
0.58
0.60
0.59
RGB
0.58
0.56
0.57
RBG
0.58 0.58 0.58
Weighted
Avg.
Table 4: SVM of Exp. 1 (“Fp”)
Precision
Recall
F-Measure
Class (QP)
0.88
0.70
0.78
QP_ref
0.56
0.50
0.53
QP_20
0.47
0.80
0.59
QP_30
0.83
0.50
0.63
QP_40
1
1
1
QP_51
0.75 0.70 0.70
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
0.50
0.72
0.59
RGB
0.50
0.28
0.36
RBG
0.50 0.50 0.48
Weighted
Avg.
Table 5: Correctly Classified Percentage (Exp. 1)
Class
(Exp. 1)
Total
Number of
Instances
Correctly
Classified
Instances
Percentage
QP (“Pv”)
50
27
54%
QP (“Fp”)
50
35
70%
C (“Pv”)
50
29
58%
C (“Fp”)
50
29
58%
Table 6: SVM of Exp. 2 (“Pv”)
Precision
Recall
F-Measure
Class (QP)
1
0.30
0.46
QP_ref
0.44
1
0.61
QP_20
0.83
0.50
0.63
QP_30
1
0.80
0.89
QP_40
1
1
1
QP_51
0.85 0.72 0.72
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
1
0.12
0.21
GRB
0.53
1
0.69
GBR
0.77 0.56 0.45
Weighted
Avg.
Table 7: SVM of Exp. 2 (“Fp”)
Precision
Recall
F-Measure
Class (QP)
0.67
1
0.80
QP_ref
0.60
0.30
0.40
QP_20
0.78
0.70
0.74
QP_30
0.91
1
0.95
QP_40
1
1
1
QP_51
0.79 0.80 0.78
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
0.86
0.24
0.38
GRB
0.56
0.96
0.71
GBR
0.71 0.60 0.54
Weighted
Avg.
Table 8: Correctly Classified Percentage (Exp. 2)
Class
(Exp. 2)
Total
Number of
Instances
Correctly
Classified
Instances
Percentage
QP (“Pv”)
50
36
72%
QP (Fp”)
50
40
80%
C (“Pv”)
50
28
56%
C (“Fp”)
50
30
60%
Table 9: SVM of Exp. 3 ("Pv")
Precision
Recall
F-Measure
Class (QP)
0
0
0
QP_ref
0.39
0.70
0.50
QP_20
0.54
0.70
0.61
QP_30
1
0.90
0.95
QP_40
1
1
1
QP_51
0.59 0.66 0.61
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
0.62
0.64
0.63
BRG
0.63
0.60
0.61
BGR
0.62 0.62 0.62
Weighted
Avg.
Table 10: SVM of Exp. 3 ("Fp")
Precision
Recall
F-Measure
Class (QP)
0.67
0.80
0.73
QP_ref
0.56
0.50
0.53
QP_20
0.80
0.80
0.80
QP_30
0.90
0.90
0.90
QP_40
1
0.90
0.95
QP_51
0.78 0.78 0.78
Weighted
Avg.
Precision
Recall
F-Measure
Class (C)
0.50
0.64
0.56
BRG
0.50
0.36
0.42
BGR
0.50 0.50 0.49
Weighted
Avg.
Table 11: Correctly Classified Percentage (Exp. 3)
Class
(Exp. 3)
Total
Number of
Instances
Correctly
Classified
Instances
Percentage
QP (“Pv”)
50
33
66%
QP (“Fp”)
50
39
78%
C (“Pv”)
50
31
62%
C (“Fp”)
50
25
50%
SMO algorithm, as indicated in Tables 5 (Exp. 1), 8 (Exp.
2), and 11 (Exp. 3).
Tables 3 and 4 display many values in boldface
font in the “Pv” case, such as “QP_ref,” “QP_40,” and
“QP_51,” regardless of whether coded degradation
occurs. On the other hand, in the “Fp” case, values in
boldface font appear for all, with the exception of
“QP_20.” For Class “C,” the “Recall” of “RGB” is
greater than 0.7. Table 5 indicates that, because the
correctly classified percentage for QP (“Fp”) is 70, we
can determine this as “classified.” For other Classes, we
can determine “not classified” because the percentage is
higher than 50.
Tables 6 and 7 indicate that in the “Pv” case,
“Precision” is greater than 0.7, with the exception of
“QP_20.” In the “Fp” case, all the items for “QP_30,”
“QP_40,” and “QP_51” are greater than 0.7. For Class
“C,” the “Precision” of “GRB” and the “Recall” and “F-
Measure” of “GBR” are greater than 0.7. Table 8
indicates that, because the correctly classified percentage
for QP (“Pv”) is less 70, and for QP (“Fp”) it is 80, we
can determine both as “classified.” Because C (“Pv”) and
C (“Fp”) are between 50% and 60%, we can determine
these as “not classified.”
Tables 9 and 10 indicate that in the “Pv” case,
“Recall” is greater than 0.7, with the exception of
“QP_ref.” In the “Fp” case, similar to Exp. 2, all t he items
for “QP_30,” “QP_40,” and “QP_51” are greater than 0.7.
For Class “C,” all the items for both “Pv” and “Fp” are
less than 0.7. Table 11 indicates that for Class “QP,”
because “Pv” and “Fp” are greater than 60% and 70%,
respectively, we can determine both as “classified,”
similar to Exp. 2. For Class “C,” because “Pv” is less than
60% and “Fp” is 50%, we can determine both as “not
classified;” however, “Pv” shows the highest correctly
classified percentage of all three experiments.
4. DISCUSSION
From the experiment results, we can see that when the
low frequency component includes B, compared with
other patterns (R and G), and coded degradation does not
occur, the assessment value is higher. However, when
coded degradation occurs, the assessment value tends to
decline rapidly. We consider that “Pv,” as one of the
factors where such tendency is seen, is nearly a blue color,
for example, the sea or sky; therefore, it is not easy to
perceive watermarking when the low frequency
component includes B. From this, on the other hand, we
consider that the assessor focuses on the perception of
coded degradation. From the SVM results, we can see
that the “Precision,” “Recall,” and “F-Measure” for “QP”
in Exp. 3 are the highest of all experiments. These
analysis results can be considered one of basis. In the case
where we watermarked the background of the CG content,
the luster and texture of the object region possibly
affected the evaluation. However, in this experiment, no
relation to the background, there is a tendency for easily
perceiving when the low frequency component includes
a green or red system. This can also be understood from
the evaluation results. We consider that this can be
applied to content creation or display and presentation
methods in the near future. In this study, we performed
watermarking by a three-level frequency domain. In the
case where the method does not divide the color
component, the assessment value is shown to be
comparatively higher. However, as is the case with this
study, when the method divides the color component, it
is clear from the experiment results that the assessment
value is affected.
5. CONCLUSION
In this paper, first, we performed wavelet transformation
for 3D CG images encoded and decoded by H.265/HEVC
with 8 Viewpoints Parallax Barrier Method. Next, we
embedded the watermarking arrangement of the RGB
color information, and we performed a quality evaluation
experiment for the generated image. Subsequently, we
analyzed the results statistically, and we classified the
results for the coded image quality and color information
using SVM.
From the experiment results, as described for
Exp. 3, we determined that when the low frequency
domain includes the B component, compared with other
components (R and G), the assessment value tends to be
higher. On the other hand, when degradation occurs, the
difference is greater. We consider that we need to pay
attention to the case of content creation or presentation.
In the near future, we consider including other frequency
domains or color space in order to further improve this
study.
6. REFERENCES
[1]. K. Matsui, “Digital Watermarking Techniques and
their Comparative Indexes,” IPSJ Tech. Rep., EIP,
vol. 2, no. 14, pp. 97—104, September 1998 (in
Japanese).
[2]. N. Komatsu and K. Tanaka, “Digital
Watermarking Technology: Digital Contents
Security,” published by IIEEJ, 2004 (in Japanese).
[3]. Y. Horita, J. Ohnishi, K. Aso, and T. Murai,
Analysis of the picture quality degradation by
digital watermarking,” IPSJ Tech. Rep., AVM, vol.
32, no. 14, pp. 77—82, 2001 (in Japanese).
[4]. C. Yara and T. Yamashiro, “Digital Watermarking
of Embedding Color Image Using Wavelet
Transform,” Bulletin of Univ. of Ryukyu
repository, May 2008 (in Japanese).
[5]. N. Inoue, A. Miyazaki, M. Shimazu, and T.
Katsura, “A Method of Digital Watermark for
Image Signals Using the Wavelet Transform,”
Journal of ITE, vol. 52, no. 12, pp. 1832—1839,
December 1998 (in Japanese).
[6]. A. Miyazaki, “On the Evaluation of Performance
of Wavelet-based Image Watermarking,” IEICE
Tech. Rep., CAS2004-13, VLD2004-24, SIP2004-
27, pp. 19—24, June 2004 (in Japanese).
[7]. M. Nishio, “Studies on Digital Watermark
Embedding Intensity Against Image Processing
and Image Deterioration,” Japanese Journal of
Radiological Technology (JSRT), vol. 60, no. 4,
pp. 500—506, April 2004 (in Japanese).
[8]. N. Kawabata and M. Miyao, “3D CG Image Quality
Evaluation for the Number of Viewpoint Including
Visible Digital Watermarking with 8 Viewpoints
Glassless 3D Method,” IEICE Tech. Rep., vol. 114,
no. 386, IMQ2014-23, pp. 39—44, Nagoya Univ.,
December 2014 (in Japanese).
[9]. N. Kawabata and M. Miyao, “3D CG Image Quality
Evaluation Including the Arrangement of Visible
Digital Watermarking by Frequency Domain
Transformation with 8 Viewpoints Glassless 3D
Method,” IEICE Tech. Rep., vol. 114, no. 485,
IMQ2014-49, IE2014-110, MVE2014-97, pp.
107—112, Seikei Univ., March 2015 (in Japanese).
[10]. NICT, http://www.nict.go.jp/, accessed January 29,
2016.
[11]. IHC Evaluation Image,
http://www.ieice.org/iss/emm/ihc/image/image.php
, accessed January 29, 2016.
[12]. Autodesk, http://www.autodesk.co.jp/, accessed
January 29, 2016.
[13]. F. Bossen, D. Flynn, and Karsten Sühring, “HM
Software Manual,”
https://hevc.hhi.fraunhofer.de/svn/svn_HEVCSoft
ware/, accessed January 29, 2016.
[14]. ITU-R BT.500-13, http://www.itu.int/rec/R-REC-
BT.500, accessed January 29, 2016.
[15]. Newsight Japan, http://newsightjapan.jp/, accessed
January 29, 2016.
[16]. N. Kawabata and M. Miyao, 3D CG Image
Quality Metrics by Regions with 8 Viewpoints
Parallax Barrier Method,IEICE Trans. on
Fundamentals, Vol. E98-A, No. 08, pp. 1696—
1708, August 2015.
[17]. N. Kawabata and Y. Horita, “Statistical Analysis
and Consideration of Subjective Evaluation of 3D
CG Images with 8 Viewpoints Lenticular Lens
Method,” IMQA2013, T1-2, pp. 2332, Sept. 12-
13, 2013.
[18]. T. Mitsuhashi, T. Hatada, and S. Yano, “Image and
Visual Information Science”, ITE, CORONA
PUBLISHING CO., LTD, 2009 (in Japanese).
[19]. N. Kawabata and M. Miyao, “Statistical Analysis of
Questionnaire Survey on the Evaluation of 3D
Video Clips,” IEICE Tech. Rep., vol. 113, no. 350,
IMQ2013-22, pp. 27—32, Shizuoka Univ.,
December 2013 (in Japanese).
[20]. H. Abe, “Data Mining Tool: Weka,” Journal of ITE,
vol. 65, no. 10, pp. 1398—1401, October 2011.
[21]. P. Gastaldo, R. Zunino, and J. Redi, “Supporting
visual quality assessment with machine learning,”
EURASIP Journal on Image and Video Processing,
http://jivp.eurasipjournals.com/content/2013/1/54,
15 pages, 2013.
... We carried out the multi-view 3D image quality evaluation including visible digital watermarking in the case of changing the coded degradation regi on by H.265/HEVC or the number of viewpoints [8]. And then, we evaluated including arrangement of visible digital watermarking by frequency domain transformation [9], based on color information [10]. However, these visible digital watermarking is embedded overall, and we did not consider by regions. ...
Technical Report
Previously, we studied about multi-view 3D CG image quality evaluation including visible digital watermarking by regions. As a result, it becomes clear that there was the mutual relationship between the watermarking by regions and the coded image quality for the multi-view 3D image. However, it is not clear that there is the relation between two focus points whether background region is color or gray scale. On the other hand, it is possible for assessors to focus on the object mutually by using gray scale for the background region of an image. Therefore, in case the background region is gray scale, we need to clear the relation among the gray scale background, the watermarking by regions, and the coded image quality by evaluating for the multi-view 3D CG image quality including the watermarking by regions. In this paper, first, we carried out the subjective quality assessment of 3D CG images including visible digital watermarking by regions with 8 viewpoints parallax barrier method. From results, we analyzed and classified the evaluation values statistically for the image region and the coded image quality by Support Vector Machine (SVM). On the other hand, we measured the luminance for the generated images by using S-CIELAB color space, and we also measured the color difference by using CIEDE2000. We carried out the objective assessments by these measurements, and compared to subjective assessment mutually.
Article
Full-text available
Recently, we are able to watch 3D videos or movies increasingly without glasses. However, they are various stereological and evaluation methods for multi-view 3D with no glasses for image quality, and their display methods are not unified. In this paper, we showed 3D CG images with 8 viewpoints lenticular lens method by ACR and DSIS methods, when we analyzed the results statistically with subjective evaluation. The experiment examined whether or not assessor were able to comfortable view the images by degree of camera’s interval and viewpoints, and whether or not they perceive or annoy degree of coded degradation at certain viewpoints.
Article
Full-text available
Objective metrics for visual quality assessment often base their reliability on the explicit modeling of the highly non-linear behavior of human perception; as a result, they may be complex and computationally expensive. Conversely, machine learning (ML) paradigms allow to tackle the quality assessment task from a different perspective, as the eventual goal is to mimic quality perception instead of designing an explicit model the human visual system. Several studies already proved the ability of ML-based approaches to address visual quality assessment; nevertheless, these paradigms are highly prone to overfitting, and their overall reliability may be questionable. In fact, a prerequisite for successfully using ML in modeling perceptual mechanisms is a profound understanding of the advantages and limitations that characterize learning machines. This paper illustrates and exemplifies the good practices to be followed.
Presentation
Full-text available
Recently, we are able to watch 3D videos or movies increasingly without glasses. However, they are various stereological and evaluation methods for multi-view 3D with no glasses for image quality, and their display methods are not unified. In this paper, we showed 3D CG images with 8 viewpoints lenticular lens method by ACR and DSIS methods, when we analyzed the results statistically with subjective evaluation. The experiment examined whether or not assessor were able to comfortable view the images by degree of camera’s interval and viewpoints, and whether or not they perceive or annoy degree of coded degradation at certain viewpoints.
Article
Full-text available
This paper presents a block based digital image watermarking algorithm that is dependent on the mathematical technique of singular value decomposition (SVD). Traditional SVD watermarking already exists for watermark embedding on the image as a whole. In the proposed approach, the original image is divided into blocks, and then the watermark is embedded in the singular values (SVs) of each block, separately. The watermark embedding on a blockby- block basis makes the watermark more robust to the attacks such as noise, compression, cropping and lowpass filtering as the results reveal. The watermark detection is implemented by extracting the watermark from the SVs of the watermarked blocks. Extracting the watermark from one block at least is enough to ensure the existence of the watermark.
Article
Several studies have been conducted on visual fatigue and visually induced motion sickness. However, no assessments have been conducted to compare 2D or 3D images or video clips. Although certain studies focused on determining the influences of viewing 3D images, there has been no unified assessment method or standard regarding 3D video clips. Moreover, prior research have demonstrated the assessment examples for only one 3D system. However, there have been no assessments comparing multiple 3D systems. Thus, the comfort of the viewing experience of multi-view 3D systems is yet to be clarified. Previously, the assessments on 3D video clips and multiple 3D systems have not considered the viewer’s personal attributes such as sex, corrected eyesight, and user experience. Thus, we initially conducted a questionnaire survey on 168 subjects regarding their experience with 3D video clips. Subsequently, we performed an experiment on 3D video assessments of 92 subjects with three display types, portraying the same 3D computer graphics content. Ultimately, the statistical analysis of the experimental results revealed the influence of age generations and individual characteristics on visually induced motion sickness.
Article
This paper describes a method of digitally watermarking image signals based on the wavelet transform. Information data can be embedded in the lowest frequency components of the image signals by using a controlled quantization process. The data are then extracted by using both the quantization stepsize and the mean amplitude of the lowest frequency components without having access to the original image. Numerical experiments show that this method provides a high-quality watermark that is robust against common signal processing.
Article
Many previous studies on image quality assessment of 3D still images or video clips have been conducted. In particular, it is important to know the region in which assessors are interested or on which they focus in images or video clips, as represented by the ROI (Region of Interest). For multi-view 3D images, it is obvious that there are a number of viewpoints; however, it is not clear whether assessors focus on objects or background regions. It is also not clear on what assessors focus depending on whether the background region is colored or gray scale. Furthermore, while case studies on coded degradation in 2D or binocular stereoscopic videos have been conducted, no such case studies on multi-view 3D videos exist, and therefore, no results are available for coded degradation according to the object or background region in multi-view 3D images. In addition, in the case where the background region is gray scale or not, it was not revealed that there were affection for gaze point environment of assessors and subjective image quality. In this study, we conducted experiments on the subjective evaluation of the assessor in the case of coded degradation by JPEG coding of the background or object or both in 3D CG images using an eight viewpoint parallax barrier method. Then, we analyzed the results statistically and classified the evaluation scores using an SVM.
Article
In order to apply digital watermarking to medical imaging, it is required to find a trade-off between strength of watermark embedding and deterioration of image quality. In this study, watermarks were embedded in 4 types of modality images to determine the correlation among the watermarking strength, robustness against image processing, and image deterioration due to embedding. The results demonstrated that watermarks which were embedded by the least significant bit insertion method became unable to be detected and recognized on image processing even if the watermarks were embedded with such strength that could cause image deterioration. On the other hand, watermarks embedded by the Discrete Cosine Transform were clearly detected and recognized even after image processing regardless of the embedding strength. The maximum level of embedding strength that will not affect diagnosis differed depending on the type of modality. It is expected that embedding the patient information together with the facility information as watermarks will help maintain the patient information, prevent mix-ups of the images, and identify the test performing facilities. The concurrent use of watermarking less resistant to image processing makes it possible to detect whether any image processing has been performed or not.
Conference Paper
The digital watermark technology is now drawing attention as a useful method of protecting copyrights of digital contents such as audio, image, and video. Especially, efficient image watermarking methods have been developed in the DWT (discrete wavelet transform) domain. In the wavelet-based watermarking, similar to the wavelet-based coding, the choice of wavelet filter banks generally affects the performance of watermarking methods. Hence a natural question arises for the wavelet-based watermarking: what is the best wavelet filter bank for use? In this paper, we evaluate wavelet filter banks used in watermarking theoretically from the viewpoint of the quality of watermarked image and the robustness of watermark, and then try to answer the 'best wavelet filter bank' problem for the wavelet-based watermarking.