VIDEO QUALITY ASSESSMENT USING LAPLACIAN MODELING OF MOTION VECTOR DISTRIBUTION IN MOBILE VIDEO STREAM

Signal & Image Processing : An International Journal (SIPIJ) Vol.4, No.3, June 2013
DOI : 10.5121/sipij.2013.4309 95
VIDEO QUALITY ASSESSMENT USING LAPLACIAN
MODELING OF MOTION VECTOR DISTRIBUTION IN
MOBILE VIDEO STREAM
Ho-Sung Han, Rae-Hong Park, Dong-O Kim, and Chang-Dae Jung
Department of Electronic Engineering, Sogang University, Seoul, Korea
pencils@sogang.ac.kr, rhpark@sogang.ac.kr, hyssop@sogang.ac.kr,
jcdlove@sogang.ac.kr
ABSTRACT
Video/Image quality assessment (VQA/IQA) is fundamental in various fields of video/image processing.
VQA reflects the quality of a video as most people commonly perceive. This paper proposes a reduced-
reference mobile VQA, in which one-dimensional (1-D) motion vector (MV) distributions are used as
features of videos. This paper focuses on reduction of data size using Laplacian modeling of MV
distributions because network resource is restricted in the case of mobile video. The proposed method is
more efficient than the conventional methods in view of the computation time, because the proposed quality
metric decodes MVs directly from video stream in the parsing process rather than reconstructing the
distorted video at a receiver. Moreover, in view of data size, the proposed method is efficient because a
sender transmits only 28 parameters. We adopt the Laplacian distribution for modeling 1-D MV
histograms. 1-D MV histograms accumulated over the whole video sequences are used, which is different
from the conventional methods that assess each image frame independently. For testing the similarity
between MV histogram of reference and distorted videos and for minimizing the fitting error in Laplacian
modeling process, we use the chi-square method. To show the effectiveness of our proposed method, we
compare the proposed method with the conventional methods with coded video clips, which are coded
under varying bit rate, image size, and frame rate by H.263 and H.264/AVC. Experimental results show
that the proposed method gives the performance comparable with the conventional methods, especially, the
proposed method requires much lower transmission data.
KEYWORDS
H.263, H.264/AVC, mobile video, motion vectors, reduced-reference, video quality assessment
1. INTRODUCTION
Quality metrics have been studied to assess the image and video quality. Subjective quality
assessment (QA) is effective since it reflects closely the human perception. However, subjective
QA is inefficient in views of the computation time and cost, because it requires various
experiments of real-time QA for a large number of images/videos. Thus, instead of subjective
QA, objective QA is preferred. Objective QA shows a good performance when its result is similar
to that of subjective QA. Objective QA is also needed to meet various requirements such as the
computation time and cost.
Development of objective image/video QA (IQA/VQA) coinciding with the subjective image
quality is important because people are the ultimate evaluators [1]. Thus, IQA/VQA can be
applied to image/video processing, reproduction, and compression systems [2]. In other words, in
some restricted applications, IQA/VQA might balance the tradeoff between desired quality and
limited resources. For example, QA can be applied to video compression systems such as H.263,

96
H.264/AVC, and moving picture experts group (MPEG). Generally, the peak to signal to noise
ratio (PSNR) or mean squared error (MSE) is used to control the compression ratio, which does
not reflect the human visual system (HVS). Objective QA is preferred to effectively consider the
HVS, which will enhance the compression efficiency. Practically, video coding system was
developed in [3], where a rate control technique was proposed by using a video quality metric.
Moreover, QA can be used to measure the quality of service (QoS) for service provider. For
example, service channel provider can estimate channel condition such as the amount of additive
channel noise and data loss by measuring quality of contents [4]. Although the same video (or
image) is transmitted, users are differently perceived by channel distortions. Thus, service
channel provider can charge a rate according to service quality.
For these reasons, many image quality metrics have been developed [5]. Video quality experts
group (VQEG) of International Telecommunication Union (ITU) is working for standardization
[6, 7]. VQA methods are classified into three types depending on the availability of reference
video (or image): full-reference (FR) [8–13] that uses all the information of a reference video,
reduced-reference (RR) [14–17] that utilizes features of a reference video, and no-reference (NR)
[18–20] that does not employ any information of a reference video. The availability of the
reference video is related to transmission data size. From this fact, it is easy to see that
transmission data size is an important issue in VQA. Practically, in some applications such as
mobile devices and Internet protocol television (IPTV), transmission data size is very important
because their network provides limited resources. These applications utilize data compression
(H.263, H.264/AVC, and MPEG) to reduce transmission data size. Therefore, the methodology
using data compression is of great advantage to VQA in the previously mentioned applications.
This paper especially focuses on RR VQA for mobile videos. Recently, mobile contents are
widely used in video applications such as digital multimedia broadcasting (DMB) and video
streaming by portable devices. Practically, digital mobile video should be compressed because of
the limitation of network bandwidth in mobile network. So, video compression techniques (H.263
and H.264/AVC) are used for mobile video applications. In [21], Lee and Sim proposed an
objective video quality evaluation method for digital mobile videos. They used block, edge,
gradient features to consider block artifacts caused by video codec. Generally, conventional RR
VQAs extract features from the reference video, and transmits them, which is suitable for video,
via an ancillary channel. However, to assess quality of mobile video, some conditions should be
considered. For example, the size of RR features is very restricted (for example, the bit rate of
encoded common intermediate format (CIF) is between 64 to 512 kbps and VQEG (multimedia
test plan) recommends 10 to 64 kbps for the size of RR) and video stream (not video) is
transmitted. We consider this circumstance, so we focus on reduction of transmission data size
and directly utilize video stream information. Specifically, we use motion vectors (MVs) that are
directly obtained from video stream as features of video through the parsing process. To reduce
the transmission data size, rather than directly comparing MVs of reference and distorted videos,
we compare one-dimensional (1-D) MV distributions, several parameters of which are
transmitted to a receiver. In this process, we use a Laplacian modeling of MV distribution to
reduce transmission data size without sacrificing quality. Note that MV distributions are
computed over whole frames of a video, not on each frame. The proposed method uses MV
histograms as feature information computed over all frames of a video, because overall
impression is more efficient than impression of each frame for VQA.
The rest of the paper is structured as follows. Section 2 presents an overview of QA. Section 3
presents the proposed VQA method using Laplacian modeling of MV distributions. Experimental

results with discussions are shown Section
on future works.
2. OVERVIEW OF QA
Traditionally, the quality of a video (or an image) is estimated by calculating the
squared error (MSE). Thus, the PSNR and MSE are
compressed data and reference data in the compression algorithms (video coding and image
compression). However, they are not enough to reflect the human visua
development of image processing (such as size of an image, compression technology) camera,
storage devices, transmission bandwidth
higher quality contents to visual satisfaction.
2.1. Limitation of Conventional QA
The simplest process to calculate the difference between two images is subtraction. For example,
the PSNR and MSE are based on
and the distorted image dI is defined
‫ܧܵܯ‬൫‫ܫ‬௥
, ‫ܫ‬ௗ
൯ ൌ ‫ܧ‬ ቄ൫‫ܫ‬௥
െ ‫ܫ‬ௗ
൯
ଶ
ቅ
where the operation E denotes the expectation. From the MSE, the PSNR is defined as




=
MSE
MAX
PSNR 10log10
where MAX represents the maximum value at a pixel (generally, 255 or 1). From (1) and (2), the
PSNR and MSE just calculate the pixel
limitations to assess the image quality. The limitations can be easily observed fro
which shows examples of three distorted images that
1(b) –1(d) are distorted versions
MSE values (around 57), however each of distorte
Nevertheless they have similar MSE values, each figure shows different quality.
the MSE cannot closely reflect the
important as quality issue. To deal with
(a)
results with discussions are shown Section 4. Finally, Section 5 gives conclusions and comments
Traditionally, the quality of a video (or an image) is estimated by calculating the PSNR
squared error (MSE). Thus, the PSNR and MSE are used to calculate the difference between
However, they are not enough to reflect the human visual perception due to
storage devices, transmission bandwidth, and so on. According to this development, people need
higher quality contents to visual satisfaction. Thus, research of QA is needed.
Limitation of Conventional QA
process to calculate the difference between two images is subtraction. For example,
the PSNR and MSE are based on subtraction process. The MSE between the reference image
is defined as
൯ ቅ
denotes the expectation. From the MSE, the PSNR is defined as




MSE
MAX 2
represents the maximum value at a pixel (generally, 255 or 1). From (1) and (2), the
PSNR and MSE just calculate the pixel-wise difference value of an image. So, there are
limitations to assess the image quality. The limitations can be easily observed fro
shows examples of three distorted images that give different visual perception. Fig
of the original image in Figure 1(a), all of which have the similar
, however each of distorted images is differently perceived by the HVS.
Nevertheless they have similar MSE values, each figure shows different quality. This means that
the MSE cannot closely reflect the human visual perception. This problem will be more and more
deal with this problem, there have been many studies on IQA/VQA
(b)
97
gives conclusions and comments
PSNR or mean
to calculate the difference between
l perception due to
and so on. According to this development, people need
process to calculate the difference between two images is subtraction. For example,
the reference image rI
(1)
denotes the expectation. From the MSE, the PSNR is defined as
(2)
represents the maximum value at a pixel (generally, 255 or 1). From (1) and (2), the
wise difference value of an image. So, there are
limitations to assess the image quality. The limitations can be easily observed from Figure 1,
perception. Figures
1(a), all of which have the similar
d images is differently perceived by the HVS.
This means that
This problem will be more and more
this problem, there have been many studies on IQA/VQA.

(c)
Figure 1. Original image (512×768 lighthouse2
similar MSE values. (a) Original image, distortion type: (b) Salt & pepper noise (MSE: 57.21), (c)
Gaussian noise (MSE: 57.54), (d) JPEG compression (MSE: 57.78
2.2. Types of Objective QA
FR quality metrics are effective
they have a limitation on multimedia services through network because they need
reference video. The PSNR and MSE are typical FR
computationally simple, so it is easy to implement them in
closely correlated with the HVS. Recently, edge PSNR
were proposed. The edge PSNR considers the HVS based on the PSNR, utilizing the fact the
HVS is sensitive to change of features such as
(ITU-T) Standardization Sector recommendation J.144:
using structure information of images [
QA by using the mean, variance, and covariance of reference and distorted images. Also, to
enhance the SSIM, a gradient-based structural similarity (GSSIM) [
the fact that the HVS is sensitive to
metric using singular value decomposition (SVD) [
between reference and distorted images are compared.
NR quality metrics use any information of a reference video. Susstrunk and Winkler prop
color IQA method by considering compression or transmission distortion [
artifacts such as blocking, blurring, and color bleeding, from which they assessed color image
quality. Hasler and Susstrunk proposed a
image [19]. They employed the CIELAB color space to reflect the HVS, in which the mean,
standard deviation, and distance information of a color image are used.
RR quality metrics using features of reference
FR quality metrics, so they can be practically
Figure 2 illustrates a block diagram
reference video via ancillary channel. Then, receiver analyzes features of distorted video and
received data (features of the reference video) to estimate quality of distorted video.
2, the performance of RR quality metrics is dependent on the features extracted from
reference video. Thus, many feature extraction methods have been proposed for
[22], an RR quality metric is proposed using edge histogram descriptor (EHD) of
feature information. Also, Wang and Simoncelli proposed wavelet coefficients as feature
information for QA [16]. They reduced the amount of transmission data by comparing computed
(d)
512×768 lighthouse2) and examples of three distorted images that have
(a) Original image, distortion type: (b) Salt & pepper noise (MSE: 57.21), (c)
Gaussian noise (MSE: 57.54), (d) JPEG compression (MSE: 57.78).
effective since they use all the information of a reference video. However,
multimedia services through network because they need
reference video. The PSNR and MSE are typical FR measures. These methods are
computationally simple, so it is easy to implement them in real-time. However, their result is not
HVS. Recently, edge PSNR [17] and structural similarity (SSIM) [
PSNR considers the HVS based on the PSNR, utilizing the fact the
features such as edges. It was adopted as ITU-Telecommunication
Standardization Sector recommendation J.144: Annex B. Wang et al. proposed the SSIM
using structure information of images [4]. The SSIM uses the structural difference of images
the mean, variance, and covariance of reference and distorted images. Also, to
based structural similarity (GSSIM) [9] was proposed by utilizing
the HVS is sensitive to change of edges. Shnayderman et al. proposed the FR quality
metric using singular value decomposition (SVD) [11], in which SVD results of each block
between reference and distorted images are compared.
NR quality metrics use any information of a reference video. Susstrunk and Winkler prop
method by considering compression or transmission distortion [18]. They measured
quality. Hasler and Susstrunk proposed a color image quality metric using color information of an
]. They employed the CIELAB color space to reflect the HVS, in which the mean,
standard deviation, and distance information of a color image are used.
RR quality metrics using features of reference images (videos) need less transmission data than
can be practically used as QA measures for multimedia service
diagram of general RR QA. Transmitter sends features of the
received data (features of the reference video) to estimate quality of distorted video.
quality metrics is dependent on the features extracted from
Thus, many feature extraction methods have been proposed for IQA or VQA
], an RR quality metric is proposed using edge histogram descriptor (EHD) of
nformation. Also, Wang and Simoncelli proposed wavelet coefficients as feature
]. They reduced the amount of transmission data by comparing computed
98
) and examples of three distorted images that have
(a) Original image, distortion type: (b) Salt & pepper noise (MSE: 57.21), (c)
of a reference video. However,
multimedia services through network because they need all of a
These methods are
time. However, their result is not
] and structural similarity (SSIM) [10]
PSNR considers the HVS based on the PSNR, utilizing the fact the
Telecommunication
. proposed the SSIM
The SSIM uses the structural difference of images for
the mean, variance, and covariance of reference and distorted images. Also, to
] was proposed by utilizing
proposed the FR quality
], in which SVD results of each block
NR quality metrics use any information of a reference video. Susstrunk and Winkler proposed a
]. They measured
tric using color information of an
]. They employed the CIELAB color space to reflect the HVS, in which the mean,
transmission data than
multimedia services.
Transmitter sends features of the
received data (features of the reference video) to estimate quality of distorted video. As in Figure
quality metrics is dependent on the features extracted from the
IQA or VQA. In
], an RR quality metric is proposed using edge histogram descriptor (EHD) of MPEG-7 as
nformation. Also, Wang and Simoncelli proposed wavelet coefficients as feature
]. They reduced the amount of transmission data by comparing computed

99
wavelet coefficients with the generalized Gaussian density (GGD) model of reference images. If
two parameters of the GGD model are given, we can construct the GGD model and compute the
Kullback-Leibler divergence (KLD) as a QA metric.
Figure 2. Block diagram of general RR QA.
As previously stated, FR methods are not proper to assess the quality of multimedia contents that
are transmitted via network. NR methods have merit in terms of transmission data size, however
they cannot guarantee the performance. So RR methods are used for QA of multimedia contents,
and transmission data size is an important measure in QA in network services.
This paper focuses on the development of the proposed RR quality metric. The proposed RR
quality metrics use MVs as features of reference and distorted videos, noting the fact that MVs of
the reference video are changed by distortion or noise. In the proposed algorithm, we compare the
1-D (horizontal and vertical) MV histograms of the reference video with those of the distorted
video, instead of directly comparing MVs. Note that MVs represent feature information of videos,
and data size is reduced by using two 1-D MV histograms instead of the two-dimensional MV
histograms. Practically, in [10], Wang et al. proved that MVs are important factors to assess the
video quality by showing that combining the SSIM with MVs improves the performance of the
VQA algorithm.
2.3. Mean Opinion Score (MOS)
The purposed QA is to estimate the quality score that is close to human perception. In other
words, the score data (assessed by human) is needed to evaluate the performance of a QA
algorithm. The score data is called MOS. MOS can be calculated to difference MOS (DMOS) by
applying double-stimulus continuous quality-scale (DSCQS) [21]. DMOS represents human
perception for the difference between reference and distorted videos (images). DMOS is generally
used to verify the performance of FR and RR QA algorithms.
2.4. Criteria for Performance Comparison
The performance of a QA metric can be quantitatively defined by calculating the similarity of a
QA metric value and the DMOS. In this paper, as similarity criterion, the Pearson correlation
coefficient (CC) is adopted, which is recommended by VQEG [4]. The Pearson CC of data A and
B is defined as
22
))(())((
))(())((
),(
BBAA
BBAA
BA
EE
EE
C
−−
−•−
= (3)
where the operation • denotes inner product. Let A be the QA metric value and B be the DMOS,
ReceiverTransmitter

then the absolute value of the Pearson CC can be used to represent the performance of a QA
metric. That is, higher absolute value of the Pearson CC means that the QA metric is better.
To verify stability of QA metric, we also adopt outlier ratio (OR) (recommended by VQEG). An
outlier signifies an observation that lies outside the expected distribution of a particular data set,
for example, does not fall within
represents that outlier data is rare, that is, OR is a measure of
if the performance of a QA metric varies depending on the statistical characteristics of video
(small OR value), then the QA metric is unstable.
In summary, an efficient QA metric gives a large Pearson CC and small OR
used for performance comparison of QA metrics in Section
3. PROPOSED VIDEO QUALITY
In this paper, we propose two video quality metrics using MV extraction and MV Laplacian
modeling. H.264/AVC basically estimates MV of 1/4 sub
compression efficiency. H.264/AVC uses the variable block size for MV estimation, instead of
the fixed block size as in conventional
variable block size for each mode. Specifically, five internodes are supported in H.264/AVC:
SKIP, 16×16, 16×8, 8×16, and 8×8 (known as macroblock (MB) partition). Note that 8×8 mode
can be further divided into 8×8, 8×4, 4×8, and 4×4 (known as MB
gives high compression efficiency using four MB sub
type of each mode.
Figure
3.1. Laplacian Modeling of MV
3.1.1. MV Histogram based Quality Metric (MVHQM)
An RR quality metric using 1
approximates the distribution of MVs as Laplacian distribution, in which parameter in Laplacian
the Pearson CC can be used to represent the performance of a QA
er signifies an observation that lies outside the expected distribution of a particular data set,
for example, does not fall within “2×standard deviation” from the DMOS. Thus, a small OR
represents that outlier data is rare, that is, OR is a measure of stability of the QA metric. However,
(small OR value), then the QA metric is unstable.
efficient QA metric gives a large Pearson CC and small OR. Both measures are
ison of QA metrics in Section 4.
UALITY METRIC
H.264/AVC basically estimates MV of 1/4 subpixel unit, which enhances the
H.264/AVC uses the variable block size for MV estimation, instead of
d block size as in conventional standards such as H.263 and MPEG-2. Fig
can be further divided into 8×8, 8×4, 4×8, and 4×4 (known as MB sub-partitions). H.264/AVC
gives high compression efficiency using four MB sub-partitions. Parameter t represents block
ure 3. Variable block size in H.264/AVC.
MV Distribution
Histogram based Quality Metric (MVHQM)
An RR quality metric using 1-D MV histograms of a video was proposed [14
100
the Pearson CC can be used to represent the performance of a QA
er signifies an observation that lies outside the expected distribution of a particular data set,
from the DMOS. Thus, a small OR
of the QA metric. However,
. Both measures are
pixel unit, which enhances the
H.264/AVC uses the variable block size for MV estimation, instead of
2. Figure 3 shows
partitions). H.264/AVC
represents block
14]. MVHQM

modeling is sent to a receiver for VQA.
frame-accumulated 1-D MV histograms. However, in our study, an MV extraction process is not
needed because MVs are directly decoded from a parsing
block diagram of the proposed MVHQM. Information to be transmitted is MV distribution of the
original video. Since MVs reflect feature information of a video, similarity of MVs between the
reference and distorted videos are used to assess the quality of the distorted
RR quality metrics usually compare information of each frame and accumulate the quality metric
of each frame to assess the video quality.
values and the DMOS, and finally get the
Figure 4. Block diagram of
MVHQM compares the 1-D (horizontal and vertical) MV histograms of the reference video with
those of the distorted video. It uses MVs as features of the reference and distorted videos, noting
that MVs of the reference video are changed by distortion or noise.
MV histograms of the reference and distorted videos by H.263. Since th
Figure 5(a)) is different from the original video (left in Fig
Left and right columns in Figure
distorted videos, respectively. It is noted that MV histograms are changed when the
contains distortion caused by H.263.
Therefore, the similarity test between MV histograms of the reference and
used a video quality metric. To compare MV histograms of the reference and distorted videos,
two commonly used similarity tests of histograms are
histogram intersection methods.
the difference between two MV histograms.
Transmitter
modeling is sent to a receiver for VQA. Generally, an MV extraction process is needed to obtain
D MV histograms. However, in our study, an MV extraction process is not
needed because MVs are directly decoded from a parsing process at a receiver. Figure
m of the proposed MVHQM. Information to be transmitted is MV distribution of the
Since MVs reflect feature information of a video, similarity of MVs between the
reference and distorted videos are used to assess the quality of the distorted video. Conventional
of each frame to assess the video quality. We employ a linear regression with the proposed metric
, and finally get the video quality metric.
. Block diagram of the proposed MVHQM without Laplacian modeling
(horizontal and vertical) MV histograms of the reference video with
It uses MVs as features of the reference and distorted videos, noting
that MVs of the reference video are changed by distortion or noise. Figure 5 shows exam
MV histograms of the reference and distorted videos by H.263. Since the distorted video (right in
the original video (left in Figure 5(a)), MV histograms are different.
ures 5(b) and 5(c) show MV histograms of the reference and
distorted videos, respectively. It is noted that MV histograms are changed when the
contains distortion caused by H.263.
Therefore, the similarity test between MV histograms of the reference and distorted videos can be
To compare MV histograms of the reference and distorted videos,
two commonly used similarity tests of histograms are considered: the histogram difference and
The histogram difference is a simple similarity test that
the difference between two MV histograms.
ReceiverChannel
101
Generally, an MV extraction process is needed to obtain
D MV histograms. However, in our study, an MV extraction process is not
ure 4 shows a
m of the proposed MVHQM. Information to be transmitted is MV distribution of the
Since MVs reflect feature information of a video, similarity of MVs between the
video. Conventional
We employ a linear regression with the proposed metric
without Laplacian modeling.
(horizontal and vertical) MV histograms of the reference video with
It uses MVs as features of the reference and distorted videos, noting
5 shows examples of
e distorted video (right in
5(a)), MV histograms are different.
5(b) and 5(c) show MV histograms of the reference and
distorted videos, respectively. It is noted that MV histograms are changed when the video
distorted videos can be
To compare MV histograms of the reference and distorted videos,
: the histogram difference and
m difference is a simple similarity test that computes

Figure 5. Frame-accumulated MV histograms of reference (left) and distorted video (right); (a)
1st frame of reference and distorted videos
of reference video, (c) Frame-accumulated MV histograms (
The histogram difference, MVHdiff
∑ −=
xv
d
xx
r
xdiff vhvhMVH ()(
(a)
(b)
(c)
accumulated MV histograms of reference (left) and distorted video (right); (a)
distorted videos, (b) Frame-accumulated MV histograms (
accumulated MV histograms (y direction) of distorted video.
diff is defined as
∑ −+
yv
y
d
yy
r
yx vhvh )()()
102
accumulated MV histograms of reference (left) and distorted video (right); (a)
accumulated MV histograms (x direction)
direction) of distorted video.
(4)

103
where the superscripts r and d represent the reference and distorted videos, respectively. The
subscripts x and y denote horizontal and vertical directions of MV projections, respectively. hx
and hy represent the 1-D horizontal and vertical MV histograms, respectively. Similarly, the
histogram intersection, MVH int between two MV histograms is defined as
∑∑ +=
yx v
y
d
yy
r
y
v
x
d
xx
r
xint vhvhvhvhMVH ))(),(min())(),(min( (5)
where min denotes the minimum operation that takes the smaller value of the two values. To test
the efficiency of the proposed algorithms, the proposed metrics are compared with the MOS. We
employ a linear regression with the proposed metrics and the differential MOS (DMOS). The
final motion vector histogram based quality metric (MVHQM) is defined as
kkkk cMVHaMVHQM +×= (6)
where the subscript k represents the type (diff or int) of the similarity test of MV histograms used.
Constant a and b can be computed by a linear regression between the proposed MVHQM and the
DMOS. Two metrics MVHQMint and MVHQMdiff give similar results. Thus, two video quality
metrics are called the MVHQM.
3.1.2. Laplacian Modeling of MV Distribution
Generally, MV distribution can be well assumed as having a Laplacian distribution. In [22], a
Laplacian distribution was used for measuring the motion compensation error, context-adaptive
arithmetic coding, or understanding of noise distribution. Therefore, we fit the distribution of
MVs with a Laplacian distribution, in which the Laplacian distribution is defined as







 −
−=
β
α
β
βα
v
vf exp
2
1
),( (7)
where α represents the location parameter and β signifies the scale parameter.
To obtain parameters of a Laplacian distribution, we minimize the difference between the MV
histogram and the Laplacian distribution, in which we use the chi-square distance [23] as a
histogram difference measure. The chi-square distance is defined as
∑ +
−
=
v
vhvh
vhvh
hhd
)()(
))()((
),(
21
2
21
21 (8)
where ν represents an index of distribution. Minimization to obtain Laplacian model parameters is
defined as
)),|(,(minarg,
)),|(,(minarg,
,
,
βαβα
βαβα
βα
βα
y
r
yyy
x
r
xxx
vfhd
vfhd
=
=
(9)
where the subscripts x and y represent horizontal and vertical directions of MVs, respectively.

104
MV distribution of the original video is expressed using Laplacian parameters, which are given
by
).,|(=
),|(=
yyy
m
y
xxx
m
x
βανfh
βανfh
(10)
The difference of MV distribution of the original video and modeled Laplacian distribution is
computed using the chi-square distance as
).,(=
),(=
r
y
m
y
r
y
r
x
m
x
r
x
hhdd
hhdd
(11)
Figure 6 illustrates an example of Laplacian modeling for horizontal MVs. Left and right columns
in Figures 6(a) and 6(b) show MV histograms of the reference and modeled Laplacian
distribution, respectively. In Figure 6(b), the distribution of MVs is modeled using two
parameters α and β of the Lapacian distribution. Note that α=0 is used since MVs have small
values close to zero and only β is estimated.
(a) (b)
(c) (d)
Figure 6. Example of Laplacian modeling for horizontal direction MV; (a), (b) MV histogram of
reference video and modeled Laplacian distribution (α≠0), (c), (d) MV histogram of distorted
video and modeled Laplacian distribution (α=0).

105
Laplacian results are obtained through minimization using the chi-square method. If we know
Lapalcian parameters, the MV distribution of the original video can be approximately obtained.
Through this process we can obtain four parameters: Laplacian modeling parameters of the MV
distribution in the horizontal and vertical directions ( β௫ and β௬, respectively) and the
differences of the original MV distribution and the modeled distribution in the horizontal and
vertical directions (݀௫
௥
and ݀௬
௥
, respectively). The total number of parameters are sent to a
receiver as the features, and the final VQA is computed at the receiver. Figure 7 illustrates a block
diagram of the proposed RR VQA, in which transmitter and receiver sides are separated.
At the transmitter, ݄௫
௥
and ݄௬
௥
in the horizontal and vertical directions of the original video can
be obtained through MV distribution extraction, and the previously mentioned four parameters
estimated through Laplacian modeling are transmitted to the receiver. At the receiver, MVs
obtained from the parsing process are used to construct histograms and obtain MV distribution
(݄௫
ௗ
, ݄௬
ௗ
) in the horizontal and vertical directions. The proposed VQA method using Laplacian
modeling of MV distribution (VLMVD) is an RR VQA method with the four parameters.
3.2. Proposed VLMVD
In this paper, the proposed VQA transmits the total number of four parameters. The total number
of 8-byte meta-data are sent through the ancillary channel, in which two bytes are assigned to
each parameter. Our algorithm has an advantage that it requires a smaller amount of meta-data
than MVHQM or conventional methods such as FR and RR VQA methods. ݄௫
௠
and ݄௬
௠
can be
obtained from Laplacian parameters β௫
, β௬
which are transmitted through the ancillary channel.
Figure 7. Block diagram of the proposed MVHQM with Laplacian modeling.
Therefore, obtained ݄௫
௠
and ݄௬
௠
from Laplacian distribution and MV distribution ݄௫
ௗ
and ݄௬
ௗ
obtained from distorted video conduct chi-square method to get difference among parameters, in
which ݀௫
ௗ
ሺ݄௠
, ݄ௗ
ሻ and ݀௬
ௗ
ሺ݄௠
, ݄ௗ
ሻ can be obtained in horizontal and vertical directions
.=ˆ
=ˆ
d
y
r
yy
d
x
r
xx
ddd
ddd
-
-
(12)
Channel ReceiverTransmitter

106
Thus, we can obtain difference values. The final video quality metric VLMVD is defines as







 ++
=
0
2
)ˆˆ(1
log
d
dd
VLMVD
yx
(13)
where ݀଴ is a constant (݀଴ ൌ 0.001ሻ.
4. EXPERIMENTAL RESULTS AND DISCUSSIONS
To evaluate the performance of the proposed metric and conventional video quality metrics, we
use a total of 124 mobile videos in our experiments. This section shows the efficiency of the
proposed MVHQMs by comparing the performance and transmission data size with conventional
VQAs.
4.1. Test Video Data
Generally, the efficiency of the quality metric is dependent on how quality metric is close to the
MOS. Therefore, to evaluate the performance of the proposed algorithm, MOS data (scores) for
test video clips are needed. We use the MOS data that are obtained from 30 people for test video
clips [21]. A total of 124 mobile videos are used as test video data. The test video data consist of
46 mobile videos coded with H.263 (46 clips) and H.264/AVC (78 clips).
Table 1. Test sequences coded with H.263.
Sequence
name
Format Frame
rate
Bit rate (Kbps)
Akiyo QCIF 15 28/33/60/118
Carphone QCIF
15 63/135
30 55/64/122/241
Coastguard QCIF 30 55/63/130/265
Container QCIF
15 28/33/64/120
30 32/63/128/236
Foreman QCIF
15 64/124
30 65/132/270
Hallmonitor QCIF 30 32/56/64/127/230
Mother and
daughter
QCIF 15 27/32/64/140
Stephen QCIF 30 130/259
Table tennis QCIF 30 55/64/127/240
Weather CIF 15 195
Tables 1 and 2 list the test sequences used in experiments that are coded with H.263 and
H.264/AVC, respectively, in which video format, frame rate, and bit rate are listed. The test video
clips consist of CIF/ quarter CIF (QCIF) which are coded under varying bit rates/frame rates by
H.263 and H.264/AVC. The DMOS can be computed by applying the double-stimulus continuous
quality-scale (DSCQS) [21] to the MOS.

107
4.2. Performance Comparison of VQA
As a similarity measure for VQA, we use the Pearson correlation coefficient (CC), which is one
of quality measures recommended by the VQEG [6], between the quality metric and the DMOS.
The higher the Pearson CC is, the better the video quality metric is. Additionally, we use the
outlier ratio (OR) for performance comparison of video quality metrics [6], in which the quality
metric with a lower OR is better.
The performance of the proposed method is compared with that of conventional methods; FR
VQA methods such as edge PSNR [4], video version of SSIM (VSSIM) [10], video version of the
image quality metric using singular value decomposition (VMSVD) [12], and Czenakowski
distance (CZD) [24], as well as RR VQA methods such as MVHQM [14], estimated PSNR
(EPSNR) [17], and local harmonic strength (LHS) with motion consideration factor (MCF) [25].
CZD is a downloaded software, in which only the first 25 frames are used. Table 3 shows the
computed Pearson CC values of 46 videos that are classified according to video characteristics.
Underlined values are the lowest ones among the methods compared, which means that the
performance is the worst. Generally, RR VQA methods give worse performance than FR VQA
methods because the number data used in RR VQA methods for representing original image
information is much smaller than that used in the FR VQA methods.
Table 2. Test sequences coded with H.264/AVC.
Sequence
name
Format Frame
rate
Bit rate (Kbps)
Akiyo QCIF 15 12/16/27/34/60
Carphone QCIF
15 12/16/34/62
30 24/32/59/69/135
Coastguard QCIF 30 25/34/54/64/133
Container
CIF
15
31/61/124/197
QCIF 16/32/60
Foreman
CIF
30
62/99/121/259/397
QCIF 24/32/55/62/132
Hallmonitor QCIF 30 23/32/56/62/133
Mother and
daughter
QCIF 15 12/16/30/33/65
Stephen QCIF 30 26/31/53/60/123
Table tennis
CIF
30
96/122/242/374
QCIF 24/32/58/66/123
Weather CIF 15 66/97/122/259
Figure 8 shows that edge PSNR shows the unstable performance among FR quality metrics and
that MVQHM and VLMVD show fluctuating performance among RR quality metrics. The
proposed method gives the best performance for the Akiyo sequence whereas the worst for the
Foreman sequence. The proposed method as a RR VQA method shows the same performance as
or slightly worse performance than the FR VQA method.

108
Table 3 compares the performance of VQA methods in terms of the Pearson CC according to the
video type. However, the Pearson CC should be compared over the whole video sequences for a
reliable comparison. Table 4 shows the performance comparison of the proposed and
conventional methods for the whole video sequences. Bold numbers in Table 4 represent the
methods with the best performance.
Table 3. Comparison of the Pearson CC of the proposed MVHQM and conventional methods for
different video clips
FR VQA RR VQA
Edge
PSNR
VSSIM VMSVD CZD MVHQM
Proposed
VLMVD
EPSNR LHS
Akiyo 0.9912 0.9878 0.9888 0.9966 0.9830 0.9999 0.9986 0.9710
Carphone 0.9990 0.9913 0.9875 0.9844 0.9990 0.7806 0.9944 0.9755
Coast
gruard
0.9894 0.9880 0.9807 0.9973 0.9598 0.9565 0.9993 0.9696
Container 0.6652 0.9398 0.9098 0.9581 0.8441 0.8160 0.9715 0.9314
Foreman 0.4748 0.9657 0.9644 0.9748 0.9174 0.7303 0.9741 0.9443
Hall
monitor
0.8385 0.9510 0.9196 0.9804 0.7839 0.9910 0.9925 0.9177
Mother
and
daughter
0.5702 0.9951 0.9967 0.9902 0.8060 0.9900 0.9685 0.9962
Silent 0.9543 0.9873 0.9816 0.9950 0.9446 0.9860 0.9677 0.9046
Table
tennis
0.9906 0.9882 0.9585 0.9970 0.9977 0.9985 0.9925 0.9755
Generally, FR VQA methods give better results than RR VQA methods in terms of the Pearson
CC. However, the proposed VLMVD gives almost the same performance as the FR VQA
methods. The proposed method gives the best performance in terms of the OR (the smaller, the
better). The proposed VLMVD can give more stable results than the other VQA methods. When
compared in terms of the amount of data required for FR VQA methods, a factor of 1.5 is
multiplied with the number of frames T for CIF/QCIF video with 4:2:0 YCbCr, because FR VQA
methods use both the luminance and (subsampled) chrominance channels. The proposed LHS
method with MCF (LHS only) and the proposed metric with MCF (LHS+MCF) are RR VQA
methods, with the amount of data required equal to only 1/1024. Note that FR VQA methods
need a large number of bits because they require the whole original video. Each frame of CIF and
QCIF sequences requires 100 and 25 bytes, respectively. EPSNR requires 68 and 17 bytes per
frame for CIF and QCIF sequences [17], respectively. In FR VQA methods, meta-data to be sent
increases as the number of the frames increases. However, MVHQM requires 520 bytes per
sequence regardless of the number of frames in the sequence, greatly reducing the amount of data
required compared with FR VQA methods. The proposed VLMVD requires a total of 8 bytes,
greatly reducing the meta-data that should be sent compared with FR as well as RR VQA
methods.

109
(a)
(b)
Figure 8. Pearson CC of two test sequences of the conventional methods and proposed method,
(a) Akiyo, (b) Foreman.
In summary, experiments with various test sequences show that the proposed VLMVD has an
advantage that it requires a small amount of data, which is an important factor in RR VQA
methods. Also the proposed VLMVD gives results similar to the conventional FR and RR VQA
methods even though it requires less amount of data for meta-data.
0.95
0.96
0.97
0.98
0.99
1
1.01
Edge
PSNR
VSSIM VMSVD CZD MVHQM Proposed
VLMVD
EPSNR LHS with
MCF
Akiyo
0
0.2
0.4
0.6
0.8
1
1.2
Edge
PSNR
VSSIM VMSVD CZD MVHQM Proposed
VLMVD
EPSNR LHS with
MCF
Foreman

110
Table 4. Performance comparison of the proposed MVHQMs and conventional quality
metrics in terms of the Pearson CC, outlier ratio, and data size for the sequences coded with
H.263/264/AVC
VQA Pearson CC OR Data size (byte)
FRVQA
Edge PSNR 0.5510 0.3261
352×288×T×1.5 (CIF)
176×144×T×1.5
(QCIF)
VSSIM 0.6538 0.3043
VMSVD 0.7118 0.3261
CZD (result for
first 25 frames)
0.4903 0.3043
RRVQA
MVHQM 0.7022 0.3696 520
(for a video clip)
Proposed
VLMVD
0.6817 0.2391
2×4
(for a video clip)
EPSNR 0.5938 0.2826
68×T (CIF)
17×T (QCIF)
LHS with MCF 0.7016 0.3043
100×T (CIF)
25×T (QCIF)
5. CONCLUSIONS
This paper proposes two RR algorithms; MVHQM with and without Laplacian modeling.
Conventional FR VQA or RR VQA methods are not suitable for measuring the quality of
multimedia services over the network because of a large amount of data required. The proposed
method focuses on reducing the amount of data while ensuring the performance for required
services. The proposed method approximates the MV distribution of the video using Laplacian
modeling. Experimental results with various test sequences show that the proposed method gives
the performance comparable to that of the conventional methods, while greatly reducing the
amount of data to be transmitted, which is desirable for a quality metric of multimedia services.
Another advantage is to get the MV in the parsing process without MV extraction process at the
receiver.
The proposed method requires a greatly small number of bits, which is suitable for mobile
services. Further research will focus on the extension of the proposed method to high-definition
video or video containing texture.
REFERENCES
[1] Pinson, M. H., and Wolf, S., “A new standardized method for objectively measuring video quality,”
IEEE Trans. Broadcasting, Sept. 2004, vol. 50, no. 3, pp. 312−322.
[2] Farias, M. C. Q., Carli, M., and Mitra, S. K., “Objective video quality metric based on data hiding,”
IEEE Trans. Consumer Electronics, Aug. 2005, vol. 51, no. 3, pp. 983−992.
[3] Yasakethu, S. L. P., Fernando, W. A. C., Adedoyin, S., and Kondoz, A., “A rate control technique for
off line H.264/AVC video coding using subjective quality of video,” IEEE Trans. Consumer
Electronics, Aug. 2008, vol. 54, no. 3, pp. 1465−1473.
[4] ITU-T Recommendation J.144, “Objective perceptual video quality measurement techniques for
digital cable television in the presence of a full reference,” International Telecommunication Union,
Mar. 2004.
[5] Sheikh, H. R., Wang, Z., Cormack, L., and Bovik, A. C., “LIVE image quality assessment database
release 2,” http://live.ece.utexas.edu/ research/quality.

111
[6] VQEG, “Draft final report from the video quality experts group on the validation of objective models
of multimedia quality assessment, phase I,” 2008, http://www.vqeg.org/.
[7] ITU-T Recommendation BT.500-11, “Methodology for the subjective assessment of the quality of
television pictures,” International Telecommunication Union, Jan. 2002.
[8] Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P., “Image quality assessment: From error
visibility to structural similarity,” IEEE Trans. Image Processing, Apr. 2004, vol. 13, no. 4, pp.
600−612.
[9] Chen, G.-H., Yang, C.-L., and Xie, S.-L., “Gradient-based structural similarity for image quality
assessment,” in: Proc. Int. Conf. Image Processing, Oct. 2006, Atlanta, GA, pp. 2929−2932.
[10] Wang, Z., Lu, L., and Bovik, A. C., “Video quality assessment based on structural distortion
measurement,” Signal Processing: Image Communication, Feb. 2004, vol. 19, no. 2, pp. 121−132.
[11] Shnayderman, A., Gusev, A., and Eskicioglu, A. M., “An SVD-based grayscale image quality
measure for local and global assessment,” IEEE Trans. Image Processing, Feb. 2006, vol. 15, no. 2,
pp. 422−429.
[12] Tao, P. and Eskicioglu, A. M., “Video quality assessment using M-SVD,” in: Proc. SPIE-IS&T
Electronic Imaging, Jan. 2007, vol. 6494, San Jose, CA, pp. (08)1−10.
[13] Han, H.-S., Kim, D.-O, and Park, R.-H. “Structural information-based image quality assessment using
LU factorization,” IEEE Trans. Consumer Electronics, Feb. 2009, vol. 55, no. 1, pp. 165-171.
[14] Han, H.-S., Kim, D.-O, Park, R.-H., and Sim, D.-G., “Visual quality metric using one-dimensional
histograms of motion vectors,” in: Proc. SPIE Image Quality and System Performance V, Jan./Feb.
2008, San Jose, CA, vol. 6808, pp. (52)1−8.
[15] Kim, D.-O, Park, R.-H., and Sim, D.-G., “Joint feature-based visual quality assessment,” Electron.
Lett., Oct. 2007, vol. 43, no. 21, pp. 1134−1135.
[16] Wang, Z. and Simoncelli, E. P., “Reduced-reference image quality assessment using a wavelet-
domain natural image statistic model,” in: Proc. SPIE Human Vision and Electronic Imaging X, Jan.
2005, San Jose, CA, vol. 5666, pp. 149-159.
[17] Yamada, T., Miyamoto, Y., Serizawa, M., and Harasaki, H., “Reduced-reference based video quality
metrics using representative-luminance values,” in: Proc. Int. Workshop on Video Processing and
Quality Metrics for Consumer Electronics, Scottsdale, AZ, Jan. 2007.
[18] Susstrunk, S. and Winkler, S., “Color image quality on the Internet,” in: Proc. IS&T/SPIE Electronic
Imaging 2004: Internet Imaging V, Jan. 2004, San Jose, CA, vol. 5304, pp. 118−131.
[19] Hasler, D. and Susstrunk, S., “Measuring colorfulness in natural images,” in: Proc. IS&T/SPIE
Electronic Imaging 2003: Human Vision and Electronic Imaging VIII, Jan. 2003, San Jose, CA, vol.
5007, pp. 87−95.
[20] Wang, Z., Sheikh, H. R., and Bovik, A. C., “No-reference perceptual quality assessment of JPEG
compressed images,” in: Proc. IEEE Int. Conf. Image Processing, Sept. 2002, Rochester, NY, vol. 1,
pp. 477−480.
[21] Lee, S.-O., and Sim, D.-G., “New full-reference visual quality assessment based on human visual
perception,” in: Proc. Int. Cont. Consumer Electronics, Jan. 2008, Las Vegas, NV, pp. 1−2.
[22] Kuo, T.-Y., “Fast variable size motion estimation for H.264 using likelihood and correlation of motion
field,” IEEE Trans. Circuits and Systems for Video Technology, Oct. 2006, vol. 16, no. 10, pp.
1185−1198.
[23] Brunelli, R. and Mich, O., “On the use of histograms of image retrieval,” in: Proc. IEEE Int. Conf.
Multimedia Computing and Systems, June 1999, Florence, Italy, vol. 2, pp. 143−147.
[24] http://www.semaca.co.uk/homepage.php.
[25] Gunawan, I.P. and Ghanbari, M., “Reduced-reference video quality assessment using discriminative
local harmonic strength with motion consideration,” IEEE Trans. Circuits and Systems for Video
Technology, Jan. 2008, vol. 18, no.1, pp. 71-83.

112
Authors
Ho-Sung Han received the B.S. degree in electronic engineering from Sogang University in 2006. Currently
he is working toward the M.S. degree in electronic engineering from Sogang University. His current
research interests are image/video quality assessment.
Rae-Hong Park was born in Seoul, Korea, in 1954. He received the B.S. and M.S. degrees in electronics
engineering from Seoul National University, Seoul, Korea, in 1976 and 1979, respectively, and the M.S.
and Ph.D. degrees in electrical engineering from Stanford University, Stanford, CA, in 1981 and 1984,
respectively. In 1984, he joined the faculty of the Department of Electronic Engineering, School of
Engineering, Sogang University, Seoul, Korea, where he is currently a Professor. In 1990, he spent his
sabbatical year as a Visiting Associate Professor with the Computer Vision Laboratory, Center for
Automation Research, University of Maryland at College Park. In 2001 and 2004, he spent sabbatical
semesters at Digital Media Research and Development Center, Samsung Electronics Co., Ltd. (DTV
image/video enhancement), Suwon, Korea. In 2012, he spent a sabbatical year in Digital Imaging Business
(R&D Team) and Visual Display Business (R&D office), Samsung Electronics Co., Ltd., Suwon, Korea.
His current research interests are video communication, computer vision, and pattern recognition. He
served as Editor for the Korea Institute of Telematics and Electronics (KITE) Journal of Electronics
Engineering from 1995 to 1996. Dr. Park was the recipient of a 1990 Post-Doctoral Fellowship presented
by the Korea Science and Engineering Foundation (KOSEF), the 1987 Academic Award presented by the
KITE, and the 2000 Haedong Paper Award presented by the Institute of Electronics Engineers of Korea
(IEEK), the 1997 First Sogang Academic Award, and the 1999 Professor Achievement Excellence Award
presented by Sogang University. He is a co-recipient of the Best Student Paper Award of the IEEE
International Symposium on Multimedia (ISM 2006) and IEEE International Symposium on Consumer
Electronics (ISCE 2011).
Dong-O Kim received the B.S. and M.S. degrees in electronic engineering from Sogang University, Seoul,
Korea, in 1999 and 2001, respectively. Currently, he is working toward the Ph.D. degree in electronic
engineering at Sogang University. His current research interests are image quality assessment and physics-
based computer vision for computer graphics.
Chang-Dae Jung received the B.S. degree in electronic engineering from Namseoul University, Cheonan,
Korea, in 2012. Currently he is working toward the M.S. degree in electronic engineering from Sogang
University. His current research interests are image/video quality assessment.

VIDEO QUALITY ASSESSMENT USING LAPLACIAN MODELING OF MOTION VECTOR DISTRIBUTION IN MOBILE VIDEO STREAM

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to VIDEO QUALITY ASSESSMENT USING LAPLACIAN MODELING OF MOTION VECTOR DISTRIBUTION IN MOBILE VIDEO STREAM (20)

Recently uploaded (20)

VIDEO QUALITY ASSESSMENT USING LAPLACIAN MODELING OF MOTION VECTOR DISTRIBUTION IN MOBILE VIDEO STREAM