FORWARD ERROR CORRECTION FOR VIDEO CODING

A common solution

João Tavares and Antonio Navarro

Telecommunications Institute, Aveiro University, University Campus, Aveiro, Portugal

Keywords: Wireless video, Video common

protection, Joint source-channel coding, MPEG-4, H.263

Abstract: Joint source-channel coding solutions has proved to provided better performance than dealin

g with both

coding processes separately. Such improvement is achieved by jointly minimizing at the source encoder the

channel effects. In this paper, we present an adaptive scheme for forward error protection of any video

coding standard. The channel coding rate changes according to the channel bit error rate (BER). The results

are impressive. For instance, a PSNR gain of about 16.7 dB is obtained at BER=10

-2

for “Foreman” video

sequence, encoded either by H.263 or by MPEG-4 and protected using the Common scheme in comparison

to the unprotected case. As our proposed scheme is common to all video standards, it obviously provides

some video quality degradation but still acceptable. We have assessed the quality degradation of the

Common solution in comparison to the optimal scheme of protection, which uses a finite number of channel

codes. Additionally, we propose a protection solution tailored to H.263/MPEG-4 video coding with an

average PSNR improvement of about 0.2 dB relatively to the above mentioned Common solution.

1 INTRODUCTION

Nowadays, commercially available mobile

multimedia terminals based on GSM/GPRS are not

able to provide video based real time multimedia

applications with accepted quality of service.

Although, the 3rd and 4th mobile network

generations overcome the bit rate limitation, wireless

video transmission offers other important negative

aspects namely caused by the deep fast fading

effects. The random behavior of the quality of the

received signal in different channel conditions will

cause enormous damages on the reconstructed video.

Compressed video is extremely vulnerable to

transmission errors since a single erroneous bit may

lead to a considerable number of frames to be

incorrectly decoded. This behavior channel was

expected in face of a source rate close to or even

greater to the channel capacity as stated by Claude

Shannon in 1948 (Shannon, 1948). However, in

practical narrow band-limited channels, the capacity

is low and consequently the source bit rate would be

required to be less than its entropy. Therefore, the

rate-distortion is a possible solution (Ortega and

Ramchandran, 1988) which basically consists of

discarding some source information. The source

distortion is minimized if it is discarded the less

important video source information. Besides, the

communication system performance can be further

increased by joint design of source-channel coding

(Daut and Modestino, 1983). It is sometimes worth

moving source bit rate to the channel to increase the

overall communication performance.

Video coding standards were developed for error

free t

ransmission since prediction and statistical

coding techniques have been adopted. Some papers

published in the recent years have investigated both

compatible standard and non-standard schemes

(Wang, 2000) to improve error robustness of video

coding. They can be classified as Forward Error

Correction (FEC), Error Concealment (EC),

Automatic Request (ARQ), Synchronization

Markers (SM) or a combination of them. FEC

schemes are usually part of the multiplexing

standards such as H.223 and H.323.

This paper proposes some solutions for the

roblem of joint source channel coding in the

context of FEC considering video communications

over wireless environments. In this context, Frossard

and Verscheure (Frossard and Verscheure, 2001),

and Stuhlmuller et al (Stuhlmuller et al, 2000)

proposed a solution for the joint source-channel

coding problem for MPEG-2 and H.263 video,

respectively. Gnavi et al (Gnavi et al, 2002)

optimally designed some coding parameters for

H.264 algorithm. More recently, in the case of image

transmission, Grangetto et al (Grangetto et al, 2004)

331

Navarro A. and Tavares J. (2004).

FORWARD ERROR CORRECTION FOR VIDEO CODING - A common solution .

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 331-336

DOI: 10.5220/0001401603310336

 SciTePress

also solved a code rate allocation problem and

demonstrated a solution employing SPIHT and

JPEG2000 as source codecs. In our research work,

we follow a slight different approach from the above

mentioned literature. As channel codes should be

applied dynamically according to instantaneous

channel characteristics, equal error protection

schemes may reach very close performance of

unequal error protection (Tavares, 2001, Navarro,

2002) and at the same time avoiding high complex

codecs and providing a common solution covering

several video coding standards.

The paper is organized as follows. Section 2

formulates the problem. Section 3 presents the test

conditions mainly concerning with source coding

parameters. In Section 4, we devise a protection

common solution as well as an H.263/MPEG-4

tailored solution. Finally, Section 5 concludes this

paper.

2 PROBLEM FORMULATION

Signal distortion is caused by loss coding and

lost/error transmission. Let X

i,j

and be the

original and the reconstructed macroblock (MB),

respectively, where index j denote the MB number

in image i. Both indices are integer numbers. The

total distortion is usually a function of the mean

square error (MSE) given by the series,

j,i

∑∑

+∞→

−

j,ij,i

lim (1)

where M and N are the number of MBs in each

image and the number of images in the video

sequence, respectively. Usually, the series is

converted to a summation with N equal to the

distance between two consecutive INTRA pictures.

The problem consists of minimizing (1) constrained

to a given constant or even variable total resources,

j,i

RR ≤

∑∑

(2)

where R

i,j

is the bit rate assigned to MB X

i,j

Nevertheless, (1) can be decomposed into a double

summation of two terms, one related to source

distortion and the other one to channel distortion,

∑∑∑∑

====

−+−

j,ij,i

)UX(

(3)

where U

i,j

is the MB resulted from the quantization

process. Despite the first term in (3) is well known

analytically (Joshi and Fisher, 1995), the second

term is a troublesome. Firstly, as the encoder makes

use of statistical coding, the model of channel

distortion has to take into account several error

propagation phenomena. Secondly, once

convolution channel codes are used, it is only known

analytically the bit error bounds since its correction

capability depends strongly on the source statistics.

Thirdly, the right hand side terms should also model

the recovering and concealment processed employed

at the receiver. Finally, both terms are not

completely independent as expressed in (3).

Therefore, in order to increase the accuracy, we

solved (1) through simulations.

3 TEST CONDITIONS

Two video coding standards, H.263 and MPEG-4,

were used to access the performance of our forward

error protection schemes.

The H.263 (ITU-T H.263, 1998) video coding

standard is a low bit rate oriented and hybrid block-

based algorithm where video frames can be encoded

with either INTRA-frame or INTER-frame coding

modes. The H.263 standard also includes more

advanced techniques that enable better compression

performance at low bit rates and better error

resilience. They are introduced as optional annexes

in versions 2 and 3 of the standard. All these

annexes are negotiable between terminals ensuring

that the decoder is able to cope with them. As the

H.263 baseline decoder is mandatory, we have used

it in our simulations without any optional annexes.

The MPEG-4 (ISO/IEC 14496-2, 2001) is the

first multimedia standard that combines interactivity,

natural and synthetic video, audio and computer

graphics. The video coding algorithm is extended

from the previous standards with the support for

arbitrarily shaped video objects and coding tools.

Additionally to the motion and texture information,

the shape of an object can also be encoded.

The simulations were carried out considering

four test sequences at QCIF (176*144) resolution,

“Foreman”, “Coastguard”, “News” and “Container”,

each containing 300 frames (only the “Foreman”

sequence contains 400 frames). Each sequence was

coded at 15 frames per second and the rate control

has been enabled to produce constant bit rate

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

332

Figure 1: PSNR-BER curves for sequence “Foreman”, H.263 coded.

ranging from 16 to 64 kbps. One INTRA frame/VOP

was coded every 2 seconds (30 frames). Therefore,

in (1), N=30 and M=99. In H.263 coding, GOB start

codes are used by our decoder as resynchronization

markers and are searched in case of synchronization

loss. In MPEG-4 coding, the packet length was

adjusted according to the bit rate in order to have

about 9 packets per VOP. Some of the error

resilience tools in the MPEG-4 standard such as

video packets, data partitioning, HEC and

Reversible VLC have been considered. The RVLCs

are used in the encoding process but not used in the

decoding since are decoded in only one direction,

forward direction. In MPEG-4 coding, each

sequence was coded in a single scalability layer and

objects are considered rectangular coincident with

frames. Table 1 summarizes MPEG-4 encoding

parameters.

The H.263 and MPEG-4 coded bit streams were

then channel encoded with rate-compatible

punctured convolutional (RCPC) codes obtained

from a rate 1/4 - memory 4 mother code (Hagenauer,

1988), according to the proposed schemes as

described in the following sections and transmitted

over a binary symmetric channel at a constant bit

rate of 64 kbps. As the channel code rate is

decreased the source code rate must be decreased

too in order to maintain constant the transmission bit

rate. The advantage of RCPC codes is that each code

in a family gives a different level of channel

protection and all of them can be decoded with the

same Viterbi decoder. These different levels of

protection may be obtained from a given mother

code using different puncturing tables and hence the

level of channel protection may be adapted to

channel conditions with minimal complexity. A hard

decision Viterbi decoder is used to decode the

received data. It is assumed that the Viterbi decoder

is aware of the puncturing tables and the code rate.

Also, a perfect reception of Video Object Layer

header as well as time information (Temporal

Reference in the case of H.263 and VOP modulo

time and time increment for MPEG-4) is assumed

since they can be protected by a sufficient powerful

channel code. This is perfectly possible because

these fields occupy just a few tens of bytes.

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

PSNR Y (dB)

8/8

8/9

8/10

8/12

8/14

8/16

8/18

8/20

8/22

8/24

8/28

8/32

IDEAL

Table 1: MPEG-4 encoding parameters.

Source resolution QCIF (176*144)

Source frame rate 30 Hz

Number of video objects 1

Number of layers 1 (single layer coding)

VOL frame rate 15 Hz

Target bit rate 16-64 kbps

Rate control VM5+

Bits per pixel 8

INTRA period 30 frames

B-VOPs Not used

Slice resynchronization Enabled

Resynchronization packet size Adjustable

Data partitioning Enabled

RVLC Enabled but not used

AIR Disabled

NEWPRED Disabled

FORWARD ERROR CORRECTION FOR VIDEO CODING

333

Performances are evaluated in terms of image

PSNR of the luminance (Y) component in function

of channel BER, i.e. distortion-BER curves. The

PSNR is averaged over the frames of the video test

sequence. In order to obtain more reliable results,

the PSNR results of several simulations with

different noise seeds have been averaged.

4 JOINT SOURCE-CHANNEL

CODING SOLUTIONS

Most video source reference coders are optimized to

give an error-free reconstructed best quality at a

certain source coding rate assuming that the coded

bit stream is received correctly. Likewise, the

channel coders are designed for a particular channel

and BER regardless of the source error sensitivities.

On one hand, the greater the source rate, the better

video quality is achieved. On the other hand, the

lower the channel code rate, the less errors occur and

the higher transmission rate is required. However, if

the channel has limited capacity, there must be a

tradeoff between the two rates, source and channel.

Two robust decoders, H.263 and MPEG-4, were

used to determine the PSNR-BER curves for each

channel code described in the previous section.

Following this approach, we plot the objective video

quality calculated in terms of video reconstructed

PSNR versus channel BER for all available channel

code rates, and adjusting the source code rate

accordingly in order to maintain the transmission

rate constant. Figure 1 presents the H.263 PSNR-

BER curves for 12 channel codes assuming a

constant transmission rate of 64 kbps and “Foreman”

sequence encoded by the H.263 encoder. Each

PSNR-BER curve is associated to a particular

channel code and is composed of two straight lines,

one (horizontal) at the low BER values and the other

(slant) at the high BER values and a curvature in the

middle BERs. The switching points between straight

and curvature lines are then defined as starting and

end curvature points. The curves intercept each other

at cross points. Cross points indicate the BERs

where a channel code should be changed in order to

maximize the PSNR. For instance, if the channel

code rate is 8/14, it should be changed to 8/16 at

cross point 10

-2

as the BER increases. The optimal

PSNR curve is then obtained by joining all PSNR

curves at cross points. The ideal PSNR curve passes

through all corners, i.e., the highest curvature points.

The ideal curve represents the maximum possible

PSNR and is not achievable in practice.

Nevertheless, as the number of codes approaches

infinity, the optimal curve tends to the ideal. In this

situation, unequal error protection schemes are

useless since there will be a better channel code that

can correct all transmission errors when the previous

channel code with higher code rate starts to produce

residual errors (Navarro, 2002). The lower the

number of available channel codes, the higher the

mean distance will be between the optimal and the

ideal curves. Despite being independent of the

number of channels codes to be used, the ideal curve

is dependent on the particular sequence and coding

algorithm. Since, in practice, only a finite set of code

rates is available, the performance will, of course, be

inferior to the ideal.

In this section, we firstly propose a joint source-

channel coding solution with 5 channel codes to be

used for both standards, H.263 and MPEG-4. This

solution is called H.263/MPEG-4. Secondly, we

devise a common solution which holds complete

independence and therefore can be used with any

source video coding algorithm. We evaluate both

solutions in relation to the 5-codes optimal case.

Figure 2 presents the code rates of the 5 channel

codes-optimal solution as function of the BER. The

code rate changes at BERs corresponding to the

interception (cross) points between PSNR-BER

curves as discussed before. The locations of cross

points vary with the video sequence and coding

algorithm.

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.10

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

Code rate

Coastguard

Container

News

Foreman

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.10

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

Code rate

Coastguard

Container

News

Foreman

Figure 2: Optimal code rates obtained with 5 codes for

H.263 (top) and MPEG-4 (bottom).

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

334

We averaged them and attained the

H.263/MPEG-4 solution depicted in Figure 3. Figure

3 also shows the channel code switching BERs for

the Common solution for which the switching points

(BERs) are the curvature starting points on the

PSNR-BER curves. In the Common solution the

switching points depend almost on the correction

capability of channel code. The curvature starting

points are always on the left of the cross points,

therefore the switching BERs for the Common

solution occur before those of H.263/MPEG-4

solution as can be seen in Figure 3.

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

PSNR loss (dB)

Coastguard

Container

News

Foreman

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

PSNR loss (dB)

Coastguard

Container

News

Foreman

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

1.10

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

Code rate

Common

H.263/MPEG4

Figure 3: Code rates for H.263/MPEG-4 and Common

solutions.

We now investigate which degradations both

solutions achieve in comparison to the 5-codes

optimal case. Figures 4 and 5 show the PSNR losses

for the H.263/MPEG-4 and the Common solutions,

respectively.

As expected, the PSNR loss is lower in the

former solution. The highest peak loss is 3.3 dB and

was obtained for “Container” video sequence

encoded with MPEG-4 using the Common solution

at BER about 6x10

-2

. This video sequence showed

greater loss at BERs close to 10

-1

. The mean PSNR

losses calculated over the entire BER range are 0.06

dB (Fig. 4-top), 0.07 dB (Fig. 4-bottom), 0.12 dB

(Fig. 5-top) and 0.37 dB (Fig. 5-bottom). Thus the

degradation introduced by both solutions is quite

small in comparison to the 5 codes optimal solution

and can be used for any video sequence. The

Common solution can also be used for any video

coding standard.

Figure 4: PSNR loss of the H.263/MPEG-4 solution in

relation to the 5 codes-optimal solution for H.263 (top)

and MPEG-4 (bottom).

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

PSNR loss (dB)

Coastguard

Container

News

Foreman

0.00

0.50

1.00

1.50

2.00

2.50

3.00

3.50

1.0E-06 1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01

BER

PSNR loss (dB)

Coastguard

Container

News

Foreman

Figure 5: PSNR loss of the Common solution in relation to

the 5 codes-optimal solution for H.263 (top) and MPEG-4

(bottom).

FORWARD ERROR CORRECTION FOR VIDEO CODING

335

5 CONCLUSIONS

This paper proposes two sub-optimal adaptive

solutions for video coded error protection at a

transmission rate of 64 kbps. Given a channel BER,

the solutions derive a RCPC code. One of the

solutions is a video coding algorithm independent

solution, i.e. a Common solution. The other one,

named H.263/MPEG-4, is dependent on the number

of channel codes used and was tailored to H.263 and

MPEG-4 coding standards and therefore providing

better performance than the Common solution for

both standards. These two solutions were assessed in

comparison to the optimal scheme which is

dependent on the video coding algorithm, channel

codes and video test sequence. On average, the

H.263/MPEG-4 solution with 5 channel codes

provides an improvement of 0.06 dB-H.263 / 0.3

dB-MPEG-4 and a degradation of 0.06 dB-H.263 /

0.07 dB-MPEG-4 in comparison to the Common

solution and to the optimal case, respectively. These

results show that both solutions are very close to the

5 codes optimal scheme. In the future, we consider

solving the same problem but for other channels

with memory. We also intend to extend the bit rate

range.

REFERENCES

Shannon, C.E., 1948. A mathematical theory of

communication. The Bell Sys. Tech. J., vol.27, pp.

379-423.

Ortega, A. and Ramachandran, K., 1988. Rate-distortion

methods for image and video compression. IEEE

Signal Proc. Mag. , vol. 15, pp. 23-50.

Daut, D. G. and Modestino, J. W., 1983. Two-dimensional

DPCM image transmission over fading channels.

IEEE Trans. Communications, vol.31, pp. 315 - 328.

Wang, Y., Wenger, S., Wen, J., Katsaggelos, A. K., 2000.

Error resilient video coding techniques. IEEE Signal

Processing Magazine, vol. 17, July, pp. 61-82.

Frossard, P. and Verscheure, O., 2001. Joint source/FEC

rate selection for quality-optimal MPEG-2 video

delivery. IEEE Trans. on Image Processing, vol. 10,

N. 12, pp. 1815-1825.

Stuhlmuller, K., Farber, N., Link, M., and Girod, B., 2000.

Analysis of video transmission over lossy channels.

IEEE JSAC, vol. 18, N. 6, pp. 1012-1737.

Gnavi, S. et al, 2002. Rate allocation for video

transmission over lossy correlated networks.

Electronics Letters, vol. 38, N. 20, pp. 1171-1172.

Grangetto, M., Magli, E. and Olmo, G., 2004. Ensuring

quality of service for image transmission: Hybrid loss

protection. IEEE Trans. on Image Processing, vol. 13,

N. 6, pp. 751-757.

Tavares, J. and Navarro, A., 2001. Error statistical

Analysis of H.263 Syntactical Elements. In Proc.

SPIE Vol. 4310 - Visual Communications and Image

Processing. San Jose, California-USA, pp. 382-391.

Navarro, A. and Tavares, J., 2002. Adaptive Joint H.263-

Channel Coding for Memoryless Binary Channels. In

Proc. IEEE International Symposium on Personal,

Indoor and Mobile Radio Communications. Lisbon,

Portugal.

Joshi, R.L. and Fischer, T.R., 1995. Comparison of

generalized Gaussian and Laplacian modeling in DCT

image coding. IEEE Signal Processing Letters, vol. 13,

N. 6, pp. 751-757.

ITU-T Recommendation H.263, 1998. Video coding for

low bitrate communication.

ISO/IEC 14496-2, 2001. Information Technology - Coding

of Audio-Visual Objects: Part 2 – Visual.

Hagenauer, J., 1988. Rate-Compatible punctured

convolutional codes (RCPC Codes) and their

applications. IEEE Transactions on Communications,

vol. 36, no. 4, pp. 389-400.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

336