(IJCNS) International Journal of Computer and Network Security, 1

Vol. 2, No. 4, April 2010

RD-Optimisation analysis for H.264/AVC scalable

video coding
Sinzobakwira Issa1, Abdi Risaq M. Jama2 and Othman Omar Khalifa3
1
Olympia College, School of Engineering
Persiaran Raja Chulan, 50200 Kuala Lumpur, Malaysia
Issa10issa@gmail.comgmail.com
2, 3
International Islamic University Malaysia, Department of Electrical and Computer Engineering
Jalan Gombak, Box: 53100 Kuala Lumpur, Malaysia

Abstract: The development of multimedia propagations and
applications has led to a greater expansion in the field of video
transmission over a heterogeneous media as well as iterative
delivery platforms with dedicated content requirements. It is
known that conventional video coding systems encode video
content with given bitrates adapted to a specific function or
application. As a result, conventional video coding does not meet
the fundamental requirements of the state-of-the-art flexible
digital media application. The newly technology based on
scalable video coding appears as a new modus operandi that has
the ability to satisfy the underlying requirements. In this work,
a multi-users scenario was considered for an optimum
performance between multiple streams. A rate distortion
optimized video frame dropping strategy which can be applied
on active network nodes during high traffic intensity was
developed. The concept of scalability here, come to introduce the
operability of high level of suppleness coding and decoding
systems. A base layer which can display the suitable quality of
the premium file was considered and take care of the
improvement of video quality.
Keywords: Bitrates, PSNR, bandwidth, multi-users scenario
and RDO.
1. Introduction
The past few decades, starting in the early nineties, a
remarkable development has been achieved in the field of
video compression. A lot of efforts were and still are being
exerted for compressing, storing data in digital medium and
allocation over the web.
It is very crucial to have the idea of monochrome digital
video data sequence which is a set of individual pictures
called frames happening at predetermined time increments.
This frame needs to be considered as a light intensity of two
dimensions in terms of function of variable x and y; f(x, y),
where x and y denote special coordinates and the value off at
any point (x, y) is proportional of the brightness of the frame
or the gray level at the point for monochrome. The normal
standard speed at which these frames are displayed is 30
frames per second.
This representation is called canonical representative.
However, this canonical representation has negative impact
because it needs very huge amounts of memory, resulting in
impracticality of being stored or shared on the web or to be
launched into digital channel. The fact may seem as an
amusing game when we try to illustrate how it could be
done.
The clear picture is an example of a 100 minutes movie
displayed at 30 frames per second with width of frame
640x480 pixels with each pixel taking 3 bytes of memory.
The reality shows that, for each second of the movie, the
requirement be at least 27MB of memory; as a result, the
entire movie will need almost 162GB of memory. If this
movie were stored on DVD’s, then considered the current
DVD capacity of 4.7 GB, would roughly require 35 DVD’s.
Therefore, video needs to be compressed considerably for
efficient storage and sharing over the web [1]
However, there are a lot of redundancies within the video
data that can be eliminated yielding file size reduction or
compression.
2. H.264/AVC Scalable Video Coding
2.1 Basic H.264/AVC structure
The H.264/AVC standard has a range of coding tools
contributing to its high compression performance, flexibility
and robustness. However, the performance improvements
come at a cost of significantly high computational
complexity. Therefore, encoder implementations should
make use of the available coding tools effectively to achieve
the desired compression performance with the available
processing resources.
H.264/AVC is an extremely scalable video codec,
delivering excellent quality across the entire bandwidth
spectrum, from high definition television to the video
conferencing and 3G mobile multimedia. The following can
thusly be summarized as the important differences.
• Enhanced motion prediction capability
• Use of a small block-size exact match transform
• Adaptive in-loop deblocking filter
• Enhanced entropy coding methods

Figure 1. H.264/AVC structure

2 (IJCNS) International Journal of Computer and Network Security,
2.2 Scalable Video Coding
Scalable video coding is desirable in heterogeneous and
error-prone environments for various reasons. For example,
scalable coding helps streaming servers avoid congestions in
network by allowing the server to reduce the bitrate of
bitstreams whilst still transmitting a useable bitstream.
One application for scalability is to improve error
resilience in transport systems that allow different qualities
of service.
For example, the essential information could be delivered
through a channel with high error protection. Scalability can
also be used to enable different quality representations
depending on playback devices processing power.
Devices with better processing power can decode and
display the full quality version, whereas devices having
lower processing power decode the lower quality version.
2.3 Types of SVC
There are three conventional types of scalability: temporal,
quality and spatial. Temporal scalability enables adjustment
of picture rate.
a) This is commonly carried out with either disposable
pictures or disposable sub-sequences, which are
explained later on. Picture rate adjustment is then
simply done by removing these disposable parts
from the coded sequence thus lowering the frame
rate.
b) In conventional quality scalability, also known as
SNR scalability, an enhancement layer is achieved
with pictures having finer quantizers than the
particular picture in the lower reference layer[3].
In coarse-granularity quality scalability, pictures in
enhancement layers may be used as prediction
references and therefore all the enhancement layer
pictures in a group of pictures typically have to be
disposed as a unit. In fine granularity scalability,
the use of enhancement layer pictures as prediction
sources is limited and therefore finer steps of
bitrate can be achieved compared to coarse-
granularity scalability.
c) Finally, spatial scalability is used for creation of
multi-resolution bitstreams to meet different
display requirements or constraints and is very
similar to SNR scalability [5].
A spatial enhancement layer enables recovery of coding loss
between an up-sampled version of the reconstructed layer
used as a reference by the enhancement layer and a higher
resolution version of the original picture.
3. Rate Distortion Optimization
3.1 Lagrangian multiplier method
In H.264/AVC, it is the art of the encoder to have the ability
of having the effective way of encoding a given video
sequence by selecting among numerous ranges of modes and
parameters.
The encoder targets to achieve optimum rate distortion
performance by choosing the best of modes and parameters
Vol. 2, No. 4, April 2010
of a given video. Doing this, the encoder would be looking
to minimize distortion in a sequence of particular video.
Rate-distortion Optimisation (RDO) methods used in video
compression are discussed in [6] [2], which include dynamic
programming and Lagrange optimisation methods.
A Lagrange optimisation method, which is also known as
Lagrange multiplier method, offer computationally less
complex (although sometimes sub-optimal) solutions to the
optimisation problem was proposed. Due to its less complex
nature, a specific form of the Lagrange optimisation method
has been used in rate-distortion optimisation of H.264/AVC
[10].
3.2 Constrained Optimisation Problem
The objective function within source constraints is
minimized or maximized by the constrained optimization.
In the case here of video coding standards, this issue of
constrained optimization can be considered as reducing the
amount of distortion of a given video sequence meaning to
strive looking to increase the number of bits that can be
encoded in exactly that particular coding sequence[4].
Below is the mathematical representation of the constrained
optimization unit;
Let S represent all the allowable vectors and let B an
element of S, (BЄS). The objective function is defined for all
B in S as D(B) and the constraint function R(B) is defined
for all B in S . The constrained problem can be presented as:
Given a constraint Rc, find
BЄS
Subject to
The solution (BЄS*) to the problem satisfies that R (B*) ≤
Rc and D (B*) ≤D (B) for all B
In S*, where
That is, if the solution to the problem is B*, then there is
no other B in S which satisfies the constraint Rc, that will
result in a smaller value for the objective function than D
(B*). The Lagrange multiplier theory offers a way of
solving the above constrained problem (i.e. finding B*) by
representing the problem as an unconstrained problem [3].
3.3 Major Theorem
The constrained optimisation problem was presented earlier
in previous section, equation (2). The Lagrange theory
represents the constrained problem as an unconstrained
problem as follows:
Theorem: for any λ≥0, the solution B*(λ) to the
unconstrained problem
 (IJCNS) International Journal of Computer and Network Security, 3
This is considered as solution of the constrained problem in
(1) presenting Rc = R (B* (λ)) as the constraint.
Proof of the theorem
If B* (λ) is the solution to the constrained problem (4) then:
Therefore,
If this is true for all B in S, it is true for a subset of B in S
where,
Now, for the above subset and for any λ≥0:
Therefore with the constraint Rc = R (B* (λ)), the solution
B* for the unconstrained problem is also the solution for the
constrained problem.
It should be noted that the theory does not guarantee a
solution for the constrained problem. It only states that for
any λ≥0 of the unconstrained problem, there is a
corresponding constrained problem which has the same
solution as the unconstrained problem.
3.4 Optimisation problem
Consider a macroblock, for which the encoder can encode
the macroblock using only one of the ‘K’ possible modes
given by the set m = {m1,, m2, … , mK }. Let ‘M’ (M Єm) be
the mode selected to code the macroblock. In the context of
H.264/AVC, these mode allocations could be any allowable
combination of macroblock partition modes, Quantisation
Parameters (QP), choice of reference frames etc… so that
the K possible modes will include all the possible admissible
parameter combinations for the
Macroblock
Define the objective function D(M) and constraint
function R(M) , where D(M) and R(M) are distortion and
rate of the macroblock as a result of selecting a particular
coding mode. If the rate constraint is Rc, the constraint
problem is defined as:
Find the coding mode M*,
Subject to
This may be written as an unconstrained problem using a
Lagrange multiplier:
Where the solution to (4-11), M*, would satisfy,
Vol. 2, No. 4, April 2010
The optimum coding mode M* (if one exists) can be found
by solving (14). That means, when the macroblock is coded
in mode M* it would satisfy the target rate (R (M*) =
Rc). All the other modes (if they exist) that satisfy R (M)≤ Rc
will have a higher distortion than D(M*).
The term D (M) + λ .R (M) in equation (4-9) is called the
Lagrangian rate-distortion cost. The mode that minimises
the Lagrangian rate-distortion cost for a particular λ≥0
(which satisfies the rate constraint in the constraint
problem) is selected as the solution mode for the constrained
problem.
4. Methodology
4.1 Objective video quality measurement
Objective video quality measurements are used to measure
the video quality, typically in situations where fast
(sometimes online) and repeatable measurements of the
distortion or the difference between the video under test and
a reference video are needed [7].
4.2 PSNR
The Peak Signal to Noise Ratio (PSNR) is the most
commonly used objective measure of video quality. PSNR is
measured as follows:
Where n is the bit depth and MSE is the Mean Squared
Error between corresponding pixel values of the original
image and the current image of the sequence under test. For
M × N array of pixels, MSE is given by:
Where Po (i, j) denotes a pixel from the original image
and Pi (i, j) denotes the corresponding pixel from the test
image. The parameters ‘i’ and ‘j’ point to a position in the
pixel arrays.
The MSE in itself can be a measure of distortion.
However PSNR is preferred because the log scale provides a
more realistic mapping to quality variations. Therefore,
PSNR continues to be the most commonly used objective
quality measure [5].
5. Implementation
For the objectives to be achieved, software video simulation
tool JSVM was used to implement and test the algorithms.
There are plenty and different H.264/AVC reference
software. JSVM was chosen for this research due to its
flexibility of varying parameters.
JSVM codec is commonly used to test new algorithms in
the video community. The use of this reference software
4 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 4, April 2010

enables realistic comparison of the performance of different
algorithms developed by different researchers. The source
code is mainly the same as the one used in the C
programming language [8].
6. Results analysis
slices one by one. If the link is still overloaded, the spatial
enhancement layers are dropped next in the same spirit, i.e.,
scale out the enhancement layers completely sticking only to
the base layer. The optimized SVC offers better quality than
the unoptimized SVC one

In this part of the simulation, basic parameters such as

frame rate of 30 Hz, number of frames 300 and group of
pictures 16 were taken into consideration. Set of stiff video
were used to evaluate the performance such as foreman,
garden, football, flower, Claire and Carphone. The PSNR
versus bitrates graph for various group of pictures were
studied in difference circumstances. Below are different
cases that were taken into consideration:

Initially, spatial dimensioning is represented by

QCIF and CIF, but was taken without additional progressive
refinement (PR) slices. With additional PR, the transform
coefficients are refined thus the improvement of the
reconstructive pictures’ quality. The performance clearly
proves that the PSNR varies with the quality.
Figure 4. Rate distortion optimization for scalable coding
Comparing to single-layer at higher bitrates, also, when the
outgoing capacity Rout is larger than the required incoming
rate, at 1670 kB/s, the RD-optimized single-layer coding
and unoptimized coding perform the same.
This is obvious, as at higher bitrates, the network link
will rarely overflow and very few or no video packets are
lost. However, if the outgoing rate is very small, it can be
seen that SVC strategy leads to good improvements in terms
of reconstructed video quality. Table 1 shows the
improvements obtained for individual video streams for the
outgoing link Rout = 600 Kbit/s.

Figure 2. Sequential scalable coding (Foreman)

In this case, several spatial resolution or bitrates are taken

into consideration or provided by the encoded bitstream.
The result shows that the PSNR is directly proportional to
bitrates.

Figure 5. evaluation of SVC and SLC

Table 1: comparison of different video streams

Sequences Scalable Video Coding Single layer coding

Optimized Unoptimized Optimized Unoptimized

(dB) (dB) (dB) (dB)

Garden 45.0008 38.5645 42.5682 40.2658

Foreman 34.5545 35.2564 32.5654 30.1254

Figure 3. Single Layer coding
Football 37.2356 37.0052 36.2545 36.5485

Based on Lagrangian Cost Function, if a video frame is to Flower 40.3215 39.0235 37.5468 37.6256
be sent on the outgoing link, it is first placed in the output Claire 36.2597 36.4566 31.2564 32.2564
buffer. Note that, for simplicity, we don’t consider the buffer
Carphone 41.3255 38.4552 38.2545 39.2545
limitations for the simulations in here.
If the outgoing link cannot accommodate all the video
packets, it will first drop the additional enhancement PR
 (IJCNS) International Journal of Computer and Network Security, 5
7. Recommendations
Although the video coding standards exhibit acceptable
quality-compression performance in many visual
communication applications, further improvements are
desired and more features need to be added, especially for
some specific applications. The important considerations
for video coding schemes to be used within future networks
could be bases on Compression efficiency, robustness with
respect to packet loss, adaptability to different available
bandwidths and adaptability to memory and computational
power for different clients.
Several other communication and networking issues are also
relevant, such as scalability, robustness, and interactivity.
A network with a single active node was considered, in our
simulations. This could be further enhanced to more
practical situations with a hierarchy of many active network
nodes and perform rate shaping at every node accordingly.
Different values for the Lagrangian multiplier λ could be
modeled for more stringent buffer conditions. A reasonable
value for λ can be determined in maximizing the
Lagrangian cost function, since λ is determined as a
function of buffer fullness.
The scalable video coding approach could be further
extended to MCTF based scalable video codec which
employs an open-loop architecture.
8. Conclusion
The choice of a Scalable Video Coding framework in this
context brings technical and economical advantages. Under
this framework, network elements can adapt the video
streams to the channel conditions and transport the adapted
video streams to receivers with acceptable perceptual
quality. The advantages of deploying such an adaptive
framework are that it can achieve suitable QoS for video
over wired and wireless networks, bandwidth efficiency and
fairness in sharing resources [11].
The adaptive scalable video coding technology produces
bitstreams decodable at different bitrates, requiring
different computational power and channel bitrate. In
addition, the bitstream is organized with a hierarchical
syntax that enables users to easily extract only a subpart of
the data contained in the bitstream and still being able to
decode the original input video but at a reduced spatial
resolution or frame rate. This process can be applied
recursively, that is, once a new bit stream is extracted out of
the original, it can undergo successive extractions
corresponding to always lower resolutions.
References
[1] C. S. Kannangara, I. E. G. Richardson, M Bystrom, J.
Solera, Y. Zhao, A. MacLennan & R. Cooney,
"Complexity Reduction of H.264 using Lagrange
Optimization Methods," IEE VIE 2005, Glasgow, 4~6
April, 2005.
[2] H. Kim and Y. Altunbasak, "Low-complexity
macroblock mode selection for H.264/AVC encoders,"
presented at International Conference on Image
Processing, Singapore, 2004.
Vol. 2, No. 4, April 2010
[3] K. P. Lim, "JVT -I020, Fast INTER Mode Selection."
San Diego: ISO/IEC MPEG and ITU-T VCEG Joint
Video Team, 2003.
[4] X. Li. Scalable video compression via over complete
motion compensated wavelet coding. Signal Processing:
Image Communication, special issue on
subband/wavelet interframe video coding, 19:637—651,
August 2004.
[5] S.-R. Kang, Y. Zhang, M. Dai, and D. Loguinov, \Multi-
layer active queue management and congestion control
for scalable video streaming," in Proc. IEEE ICDCS,
Tokyo, Japan, Mar. 2004, pp. 768{777}.
[7] T. Oelbaum, V. Baroncini, T. K. Tan, and C. Fenimore,
“Subjective quality assessment of the emerging
AVC/H.264 video coding standard,” International
Broadcasting Conference (IBC), Sept., 2004.
[7] R. Leung and D. Taubman. Impact of motion on the
random access efficiency of scalable compressed video.
Proc. IEEE Int. Conf. Image Processing, 3:169—172,
September 2005.
[8] R. Leung and D. Taubman. Perceptual mappings for
visual quality enhancement in scalable video
compression. Proc. IEEE Int. Conf. Image Processing,
2:65—68, September 2005.
[9] R. Leung and D. Taubman. Transform and embedded
coding techniques for maximum efficiency and random
accessibility in 3-D scalable compression. IEEE Trans.
Image Processing, 14(10):1632—1646, October 2005.
[10] R. Leung and D. Taubman. Minimizing the perceptual
impact of visual distortion in scalable wavelet
compressed video. Proc. IEEE Int. Conf. Image
Processing, October 2006.
[11] R. Leung and D. Taubman. Perceptual optimization for
scalable video compression based on visual masking
principles. IEEE Trans. Circuits Syst. Video Technol.,
submitted in 2006.
[12] T. Wedi and Y. Kashiwagi, “Subjective quality
evaluation of H.264/AVC FRExt for HD movie
content,” Joint Video Team document JVT-L033, July,
2004.
[13] ISO/IEC JTC 1/SC 29/WG 11 (MPEG), “Report of the
formal verification tests on AVC/H.264,” MPEG
document N6231, Dec., 2003 (publicly available at
http://www.chiariglione.org/mpeg/quality_tests.htm).
[14] T. Schierl, T. Stockhammer and T. Wiegand, "Mobile
Video Transmission using Scalable Video Coding
(SVC)," IEEE Trans. On Circuits and Systems for
Video Technology, Special issue on Scalable Video
Coding, scheduled June 2007.
[15] S. Wenger, Y.-K. Wang and T. Schierl, “Transport and
Signaling of SVC in IP networks,” IEEE Transactions
on Circuits and Systems for Video Technology, Special
issue on Scalable Video Coding, scheduled for: March
2007.