ieee journal of selected topics in signal...

IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 1, NO. 2, AUGUST 2007 295

Cooperative Source and Channel Coding forWireless Multimedia Communications

Hoi Yin Shutoy, Deniz Gündüz, Student Member, IEEE, Elza Erkip, Senior Member, IEEE, andYao Wang, Fellow, IEEE

Abstract—Past work on cooperative communications has in-dicated substantial improvements in channel reliability throughcooperative transmission strategies. To exploit cooperation ben-efits for multimedia transmission over slow fading channels, wepropose to jointly allocate bits among source coding, channelcoding and cooperation to minimize the expected distortion ofthe reconstructed signal at the receiver. Recognizing that, notall source bits are equally important in terms of the end-to-enddistortion, we further propose to protect the more important bitsthrough user cooperation. We compare four modes of transmissionthat differ in their compression and error protection strategies(single layer or multiple layer source coding with unequal errorprotection, with versus without cooperation). Our study includesan i.i.d. Gaussian source as well as a video source employing anH.263+ codec. We present an information theoretic analysis forthe Gaussian source to investigate the effects of the modulationscheme, bandwidth ratio (number of channel uses per sourcesample), and average link signal-to-noise ratios on the end-to-enddistortion of the four modes studied. The information theoreticobservations are validated using practical channel coding sim-ulations. Our study for video considers error propagation indecoded video due to temporal prediction and jointly optimizesa source coding parameter that controls error propagation, inaddition to bits for source coding, channel coding and cooperation.The results show that cooperation can significantly reduce theexpected end-to-end distortion for both types of source and thatlayered cooperation provides further improvements and extendsthe benefits to a wider range of channel qualities.

Index Terms—Joint source channel coding, layered sourcecoding, unequal error protection, user cooperation, wireless video.

I. INTRODUCTION

THE advance of next generation wireless communicationsystems are bringing real-time video onto wireless de-

vices. Video applications are bandwidth demanding and errorsensitive, while wireless channels are unreliable and limited inbandwidth. Furthermore, the real-time nature of many videoapplications makes retransmission impossible. The key toimproved real-time wireless video transmission therefore liesin increasing channel reliability and error resilience withoutsacrificing the bandwidth efficiency [1]. The multiple-input,multiple-output (MIMO) technology has brought improve-ments in robustness to fading by providing diversity in thespatial domain. As an alternative to multiple antenna systems

Manuscript received November 1, 2006; revised May 4, 2007. This work wassupported in part by the National Science Foundation under Grant 0430885. Theassociate editor coordinating the review of this manuscript and approving it forpublication was Dr. Adriana Dumitras.

The authors are with the Department of Electrical and Computer Engineering,Polytechnic University, Brooklyn, NY 11201 USA (e-mail: [email protected]; [email protected]; [email protected]; [email protected]).

Digital Object Identifier 10.1109/JSTSP.2007.901516

when the devices are limited to a single antenna (because ofeither the limitations on the device size, and/or the cost ofadditional antennas), user cooperation or wireless relaying isnow being vigorously researched to provide spatial diversity.In cooperative communication systems, wireless transmissioninitiated from a source node is overheard by other terminals(called relays). Instead of discarding this overheard signal, therelays process and forward the signal to the intended destina-tion, where different copies of the signals are combined forimproved reliability. This spatial diversity scheme providesadaptation to the time-varying channel state by exploiting thenetwork resources instead of limiting itself to the bottlenecksof the point-to-point channel.

Sendonaris et al. [2], [3] introduced the concept of coopera-tion among wireless terminals for spatial diversity. They showedthat user cooperation is able to effectively achieve robustnessagainst fading. Laneman et al.. [4] studied different coopera-tive protocols to formulate cooperative diversity gains in termsof physical layer outage probability. Hunter and Nosratinia [5],[6] proposed a cooperative scheme using rate compatible punc-tured convolutional (RCPC) codes and cyclic redundancy check(CRC) for error detection. Stefanov and Erkip [7] presented an-alytical and simulation results studying the diversity gains ofcooperative coding and suggested guidelines on how good co-operative codes can be designed. The fundamental principlesof cooperative transmission outlined in [2]–[7] are extendedto multiple relays and/or multiple transmitter/receiver pairs in[8]–[10].

Most prior studies focus on the physical layer and use channeloutage probability or the residual error probability after channeldecoding as a performance measure. For multimedia signals(audio, image, and video), a more relevant performance measureis the distortion of the decoded signal compared with the orig-inal source. This distortion is contributed from both the sourceencoder (due mainly to quantization) and residual transmissionerrors after channel decoding. The latter depends on channel sta-tistics, the channel encoder and decoder, the source coder errorresilience features and source decoder error concealment tech-niques, as well as the diversity method employed. There is atradeoff between the source coder quality and the channel coderreliability, and the optimal allocation of the resources amongthese two requires a joint optimization.

In this paper, we optimize bit allocation among source coding(for both compression and error-resilience), channel coding,and cooperation, subject to a total bit rate constraint. In sourcecoding, not all bits are equal in importance since the rate-dis-tortion function is in general not linear. We therefore proposeto protect the more important bits through stronger channel

1932-4553/$25.00 © 2007 IEEE

296 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 1, NO. 2, AUGUST 2007

coding and user cooperation. We call this layered-cooperation.To show the benefits of cooperation and unequal error protec-tion (UEP), we compare four communication modes: directtransmission (single-layer source and channel coding withoutcooperation), layered transmission (layered source coding andUEP through channel coding, without cooperation), coopera-tive transmission (single-layer source and channel coding withcooperation), and layered cooperation (layered source codingwith UEP, and cooperation among layers). These strategiesprovide various levels of adaptivity to the overall networkstate. We show that optimal allocation of the resources incase of layered cooperation results in the highest performanceas multiple layers provide better adaptivity to the unknownchannel coefficients at the transmitters, and it can better exploitthe network resources through channel cooperation. While ourresults here focus on a simple network of three nodes, they caneasily be extended to larger network scenarios following theresults of [8]–[10].

To gain theoretical insights into the performance limits of thesefour modes of communications, we first examine the achievableminimal distortion for an i.i.d. Gaussiansource. For this study,wemake use of the well-known rate distortion function and the suc-cessive refinability of an i.i.d. Gaussian source [19] to determinethe encoding-induced distortion at different source rates with orwithout layered source coding. In order to model residual trans-mission error,weperform an information theoretical analysis anduse outage probability, defined as the probability of the instanta-neous capacity of the fading channel being less then the desiredtransmission rate [20], which has been shown to be a tight lowerbound to the error probability in the limit of infinite length code-words [21]. This provides us an understanding of the theoreticallimits achievable by the proposed schemes without assuming anyspecific channel code. Furthermore, it also enables us to easilyobserve the effects of the modulation, various link qualities andthe bandwidth ratio (number of channel uses per source sample)on the end-to-end performance.

We also study the performance achieved by practical channelcodes for transmitting a Gaussian source by performing channelsimulations using binary phase shift keying (BPSK) modula-tion and RCPC channel codes [22]. Noticeably, comparison ofthese simulation results with the information theoretic boundsfor BPSK modulation reveals similar trends in terms of theend-to-end average distortion. In particular, our results, boththeoretical and practical, show that in a modulation-constrainedenvironment cooperation can provide significant improvementsover no-cooperation when the average channel SNR is in thelow to medium range, and that layered cooperation can extendthis benefit to a larger range of channel qualities. When themaximum modulation level is not fixed (such as Gaussiancodebooks used in the outage approach), layered cooperationalways outperforms all other transmission modes, and the gapincreases with channel SNR. Furthermore, benefits of coop-erative transmission and layered cooperation increase whenwe have more channel uses per source sample, i.e., higherbandwidth ratio.

Encouraged by our results for the Gaussian source, we extendour study to include real video sequences. We use the H.263+video codec [23], [24] with SNR scalability option for video

coding, and further examine the impact of the encoder intra-block rate, a parameter that controls the encoder error resilience.Overall, we find that benefits obtained from cooperation andlayered source coding for video follow similar trends exhibitedby the Gaussian source. For the modulation-constrained case,the particular channel SNR range in which cooperation outper-forms no-cooperation, and the range in which layered cooper-ation outperforms single-layer cooperation, however, do differ.These ranges also depend on the motion content of the under-lying sequences. Moreover, a video source is more sensitiveto the user-to-relay channel quality. A high SNR user-to-relaychannel benefits video more than Gaussian source.

The literature on joint source-channel coding for cooperativesystems is sparse. In [11]–[16], we introduced the problem ofsource-channel coding in cooperative relay systems. We consid-ered single and two-layer source coders and through simulationstudies, compared achievable minimal distortions by differentcommunication modes, both for the i.i.d. Gaussian source [11],[12], [15] and a real video source [16]. In [13] and [14], westudied the high SNR behavior of end-to-end distortion, and pro-vided cooperative source and channel coding strategies that uselayered compression with progressive or simultaneous trans-mission. Results in [13], [14] are significant in the sense that,in the high SNR regime, our cooperative joint source-channelcoding schemes provide optimal decay rate of end-to-end distor-tion (known as distortion exponent [17]) with increasing SNR,for a wide range of bandwidth ratios. The authors in [18] com-bined multiple description source coding with user cooperation,and studied achievable performances when a Gaussian sourceis coded into either one or two descriptions, and sent with orwithout cooperation.

The remaining of the paper is organized as follows. InSection II, we describe the channel model, formulate thefour modes of communication and introduce the optimal bitallocation problem. In Section III, we consider transmissionof an i.i.d. Gaussian source, and analyze both the informationtheoretic limits and the practical channel coding performancefor various link qualities. In Section IV, we analyze the simu-lation results for various video sequences using H.263+ codec.Section V concludes this study.

II. SYSTEM SETUP AND FOUR MODES OF COOPERATION

We consider two terminals, and , in the same wirelessnetwork. We assume sends its source, , to its destination

and is ’s cooperative relay. As transmits towards, overhears ’s signal and relays it to . This setup

can be applied to several scenarios, provided that can over-hear ’s transmission; and can receive and combine the twosignals. For example, could be a 3G mobile phone sendingreal-time video to its base station, ; and is another mobiledevice in the network. Or, could be a wireless LAN clientreceiving streaming video from a server, then would be theaccess point and , another WLAN client who streams videofrom either for his own usage or to help receive a betterquality video. As customary in the literature [4] and to capturethe benefits of cooperation, we assume time division multipleaccess among and . While cooperating, only transmits

SHUTOY et al.: COOPERATIVE SOURCE AND CHANNEL CODING 297

during a certain part of its time slot (or channel frame); the re-maining part of the time slot for is used by to send ’sinformation. Similar to [5], [7] if cannot decode ’s signal(which we can check using an error detection mechanism suchas CRC), uses the remaining time to continue transmissionto . Note that, this requires a single bit feedback from toboth and , which can be supported by a low rate high reli-ability feedback link. This feedback rate is negligible comparedwith the data rate used in communication. Furthermore, CRCis already present in communication systems and the overheaddue to CRC is small. In general, may also be sending its ownsource to either the same or a different destination in the timeslotbelonging to ; and may or may not help in ’s transmis-sion, but in this paper, we are only interested in investigating theperformance from ’s perspective and hence, we concentrateonly on the time slot allocated for ’s transmission. Also wedo not make use of the correlation between the sources ofand , even when such correlation exists.

We assume flat, slow Rayleigh fading, with independentfading amplitudes , and for the three wireless linksas shown in Fig. 1. We define one channel frame as a blockof channel uses and assume that the channel coherencetime (over which the fading amplitude is constant) spans aninteger number of channel frames. The fading amplitudes varyindependently from one coherence time to the next. The fadingamplitude squares, , and , each follow anexponential distribution with unit mean. The additive noise isi.i.d. white Gaussian and independent for each receiver. Weuse , , 2, to denote the average received signal tonoise ratio at the destination, for channels 1 ( to ) and 2( to ), respectively. We will call them user channel SNRs.The average received SNR for the channel between and ,the inter-user channel, is . We assume that averagereceived SNRs, which model the pathloss and/or shadowing inthe environment, are known by all terminals, but instantaneouschannel realizations can only be measured at the correspondingreceivers, which are then used for maximum likelihood channeldecoding. In our analysis and simulations, we consider both anasymmetric scenario, where ,i.e., has better channels to both and than thedirect link on the average, and a symmetric scenario, where

, i.e., both terminals have the same averagequality channels to the destination. Asymmetric scenario couldhappen, for example, if is closer to than orlink is severely shadowed; symmetric scenario takes placewhen and are close to each other.

In order to describe the four modes of transmission, we con-sider a system with a fixed -ary modulation with which a totalof bits can be transmitted over the channel in one channelframe, that is, . We assume we have sourcesamples to be transmitted in one channel frame, leading to abandwidth ratio, . The bits in a channel frame areallocated among source coding (number of compressed bits rep-resenting source samples), channel coding (channel parity bitssent by the user), and cooperation (channel parity bits sent bythe relay). Modes 1 and 2 below are well-known communica-tion schemes neither of which utilizes user cooperation. Mode2 differs from mode 1 in that it uses the popular UEP method to

Fig. 1. Cooperative system model. Here T wishes to communicate with Dwith the help of the relay T . In the figure, SNR , SNR , and SNRdenote average received SNRs; h , h and h denote instantaneous fadingamplitudes.

Fig. 2. Four modes of communication described in Section II.

increase source error resilience. Mode 3 uses cooperative coding[5], [7], while mode 4 further applies cooperation to implementUEP. These communication modes are detailed in Fig. 2.

Mode 1 (Direct Transmission): Terminal compressesat source bits/frame and applies a rate channelcode, where is the number of parity bits/channel frame. Ittransmits the resulting bits directly to destination .The constraint is .

Mode 2 (Layered Source Coding and Transmission): Thisis the conventional layered source coding with UEP throughchannel coding. is compressed into two layers: a base-layer(BL) with source bits and an enhancement-layer (EL) with

source bits in each channel frame. applies a channel codeof rate for BL and one of rate forEL. All bits are sent directly to . The con-straint is . When and ,mode 2 operates as mode 1.

Mode 3 (Cooperative Channel Coding): Terminal com-presses into a single layer using bits, and then implementsa rate channel code that can be punctured torate . source bits and parity bits are thentransmitted directly to . As overhears and decodes ’stransmission, it generates and sends the remaining channelparity bits (the “cooperation bits”) for . If cannot decodethe source bits, it sends 1-bit “frame error” signal to and

. will then continue to send the parity bits itself. Thebit allocation constraint is . Notice thatwhen , mode 3 operates as mode 1.

Mode 4 (Layered Cooperation): is compressed withbits for the BL and bits for the EL. We apply cooperationonly for BL. BL is protected by parity bits sent by andcooperative parity bits sent by if can decode the base layerafter it observes channel coded bits. Otherwise, theparity bits are sent by . EL is protected by parity bits sent by


. The bit allocation constraint is .Mode 3 is a special case of mode 4 when . So,mode 4 includes both mode 3 and mode 1. Mode 2 is also aspecial case of mode 4 when . Hence, mode 4 is themost general mode under consideration.

For each of these four modes, there is an optimal bit alloca-tion, subject to a constraint on the total number of bits/channelframe , among source compression (BL and EL), channelcoding and cooperation that will lead to the lowest end-to-endsource distortion. This optimal allocation depends on the type ofthe source, the modulation, the frame size, the bandwidth ratioand the source and channel coding strategies as well as the av-erage channel SNRs. In the next few sections, we will studythis optimal bit allocation problem and compare the resultingend-to-end distortion for different sources and channel qualitiesas well as modulation modes and bandwidth ratios.

III. COOPERATION FOR AN I.I.D. GAUSSIAN SOURCE

To gain insights into the achievable performance of thevarious communication modes in Section II, we first eval-uate their performances with an i.i.d. real Gaussian sourcewith unit variance. For a bandwidth ratio of , if each sourcesample is compressed to bits, is constrained to be

. Using the distortion-rate function ofa zero mean, unit variance i.i.d. real Gaussian source [25], thedistortion per sample in terms of mean squared error (MSE) is

.Note that, this distortion-rate function is an upper bound

for the distortion achievable for any unit variance memorylesssource at the same compression rate and is asymptotically tightas goes to infinity. For sources with memory, including video,the precise form of the preceding function is not applicable, butthe exponential decay of distortion with rate is generally stilltrue [26], [27].

A. Formulation of the Expected Distortion and the OptimalBit Allocation Problem

For a given communication mode, modulation level, band-width ratio and average channel SNRs, the expected distortion(ED) is a function of source rate ( and ), the amount ofchannel coding ( and ) and the level of cooperation .Below we will discuss how to compute ED for mode 4. The for-mulation of ED for other modes can be easily derived since theyare special cases of mode 4. When the channel decoder cannotcorrect all the bit errors that occur in a channel frame for eitherBL or EL, we assume all source bits corresponding to that layerare dropped. For a given channel frame, there are three mutuallyexclusive events after channel decoding.

a) Both BL and EL are correctly decoded.b) Only BL bits are decoded correctly.c) BL is not decodable. Since without the BL, the EL cannot

be used, we simply replace all the samples by their meanvalues.

We denote the average probabilities of cases a), b), and c)(averaged over all fading realizations) as , and respec-tively. Due to the successive refinability of the Gaussian source[19], for case a), the distortion per sample is , where

and . For case b), it is and forcase c), it is 1 due to unit variance assumption. Theexpected distortion (ED) for a given channel frame is

(1)Since the source is memoryless and we assume that source

coding and decoding in each channel frame only depends onsamples in the current frame, there is no decoding error propa-gation from one channel frame to the next in the decoded sam-ples. Note that, as discussed in detail in Section IV-A, becauseof error propagation, (1) does not hold for video sources.

Before we proceed to discuss the formulation of probabilities, and , we first define the following that describe the

residual frame error rates (FER) after channel decoding.• is the average FER of the inter-user channel, averaged

over different fading levels . It can be expressed as

where is the probability of incor-rectly decoding a channel frame with source bits andparity bits when it is sent over a channel with average SNR,

, for an instantaneous fading amplitude . Gen-erally, depends on other channel coding parameterssuch as the channel code used, the modulation level andpacket size. Here, we assume these parameters are fixedand we do not explicitly show that dependence.

• is the FER of the BL frame at for a particularfading level given that the relay cannot decode the

source bits and thus cooperation does not take place.When BL is transmitted directly by , the source bitsand parity bits in a channel frame are all subject tothe fading and the average channel SNR, .can be written as

• is the instantaneous FER of the EL frame at ,which is transmitted entirely by . The source bits and

parity bits are sent over a channel with fading amplitude,and average SNR,

• is the FER of BL when cooperation takes place.It is the probability of decoding BL incorrectly whensource bits and parity bits are sent by ; whileparity bits are sent by after decoding the source bits.Thus bits are sent under a channel realization of

with average SNR, while bits are sent underwith average SNR,

For each of the cases a), b), and c), there are two mutually ex-clusive situations depending on whether is transmitted with

1Although it is possible to perform error concealment of samples in the cur-rent frame from reconstructed samples in previous frames for general sources,we do not consider this here since the source is memoryless.


or without cooperation. We define the probability of the six re-sulting outcomes as follows.

Pr(receive BL and EL, with coop.).

(2)

Pr(receive BL and EL, without coop.)

(3)

Pr(receive BL only, with coop.)

(4)

Pr(receive BL only, without coop.)

(5)

Pr(do not receive BL, with coop.)

(6)

Pr(do not receive BL, without coop.)

(7)

The probabilities , for , 1, 2, in (1) can therefore beexpressed in terms of (2)–(7) as

(8)

For a given bandwidth ratio, modulation level, channel codeand packet size, the goal is to find the optimal bit allocation

that results in the minimum ED for each, and under consideration. We also

wish to study conditions under which each of the four trans-mission modes becomes dominant and investigate the choice ofsystem parameters on the end-to-end distortion.

B. Information Theoretic Formulation of Expected Distortion

In this section, we introduce an information theoretic frame-work to solve the expected distortion optimization problem de-scribed above. In order to understand the potentials of the sug-gested four modes of communication, we will consider outageprobability which serves as a tight bound for the frame errorrate of any practical channel coding scheme in the limit of infin-itely long codewords [21]. The outage probability is defined asthe probability that the instantaneous channel capacity is lowerthan the desired transmission rate. Intuitively, the best possiblechannel code (in Shannon sense) only results in errors when wetransmit at a rate higher than the channel capacity, hence errorsin a fading environment will only occur if the desired rate ishigher than the instantaneous capacity.

For an -ary modulation scheme with soft decision de-coding, we can model a slow fading channel of the formdescribed in Section II as an -input, real-output channel. Wedenote the instantaneous capacity of this channel at fading level

, and average signal-to-noise ratio SNR as .When the modulation scheme is not constrained and weare allowed to use real valued inputs (as in Gaussian code-books), we have .Then the outage probability for rate becomes

. Note that repre-sents the desired number of information bits (or source bits) perchannel use that is to be transmitted over the fading channel.For example, for our proposed mode 1, we have

(9)

The corresponding source coding rate in terms of bits per sourcesample becomes , where is the bandwidthratio.

We now rewrite the various frame error probabilities and theexpected end-to-end distortion expression in Section III-A formode 4, in terms of the outage probabilities. For the base layer,we transmit a total of bits through the channel.For -ary modulation, this results inchannel uses. Therefore, the base layer utilizes proportion ofthe channel uses, where

Of these channel uses, the original user gets proportion,where

and the relay uses the remaining proportion. Similarly,the enhancement layer utilizes proportionof the channel uses.

The base layer is transmitted atinformation bits per channel use and the enhancement layer

at information bits per channel usetowards the destination. The transmission rate of the base layerfrom to becomes in-formation bits per channel use. The corresponding source com-pression rates for the base and the enhancement layer are

and bits per source sample.Combining, the outage probabilities corresponding to the

FERs in (2)–(7) are

(10)

(11)

(12)

(13)

(14)

(15)

with . In the above expressions,, , and correspond to, respectively, the capac-

ities of , , and links and depend onand for . Also note that, to model the

additional parity bits transmitted by when cooperation takesplace (as in ), we use independent codebooks at forretransmission of the decoded information [6]. Hence, the totalmutual information is the sum of the mutual information terms


from the source and the relay to the destination. This is higherthan the rate achievable by repetition based-scheme, which usethe same codebook at both the source and the relay ([4]).

The overall outage based end-to-end distortion can be foundusing the outage probability expressions in (10)–(15) as

(16)

with for .For fixed , , and link SNRs, the end-to-end distortion can

now be minimized over all choices of with, . Note that and are both upper bounded

by and the assumption of large channel frame lengthsused in the information theoretic approach enables us to use anyreal valued . As special cases of mode 4, if ,

, we would reduce to mode 1, if we wouldreduce to mode 2, and if , , we wouldreduce to mode 3.

It is possible to use the exact expressions for the outage prob-abilities and the distortion-rate functions in (16), and obtain anexpected distortion expression in terms of the rates , andthe channel allocation variables and . The optimal allocationof these variables is a nonlinear optimization problem whoseefficient solution is not obvious. Here, we will use exhaustivesearch over a discretized search space to obtain numerical re-sults.

C. Results for the Information Theoretic Analysis

In this subsection, we present our information theoretical re-sults, which provide an expected distortion lower bound for anypractical coding scheme that can be used. For easier comparisonwith video simulation results presented in Section IV, instead ofED we use the decoded signal SNR, , defined below, asour performance measure

In general, when there are no constraints on the max-imum modulation level, the optimal transmission scheme thatachieves the Gaussian channel capacity

requires Gaussian codebooks. We considertwo bandwidth ratios and , and study an asymmetricscenario where with Gaussiancodebooks. The optimal end-to-end signal SNR as afunction of average channel SNR, can be seen in Fig. 3for the four modes of transmission. The four topmost curves inthe figure correspond to case while the lowest four curvesare for . We observe a significant improvement in thereconstructed signal SNR, with cooperation (mode 4 and mode3). While source layering (mode 2) improves the end-to-enddistortion compared to direct transmission (mode 1), coopera-tion results in a larger gain with layered cooperation (mode 4)resulting in the best signal quality. Also note that the differencebetween the modes increase as channel SNR increases. We alsosee in Fig. 3 that the amount of improvement that modes 2, 3,and 4 bring over mode 1 increases with the bandwidth ratio,which yields more channel uses per sample. In fact, the rate of

Fig. 3. Signal SNR, SNR (dB) versus average channel SNR for T � D

link, SNR (dB) for different modes of communication with no constraints onmodulation. Results are based on information theoretic analysis. The topmost4 curves are for a bandwidth ratio of b = 4, and the 4 curves below are for abandwidth ratio of b = 1. We assume an asymmetric scenario with SNR =

SNR = 4SNR .

increase of with channel SNR, gets larger with the band-width ratio and mode 4 consistently results in the largest rate ofincrease. This is consistent with our analysis of the distortionexponent of layered cooperation in [13], [14] where we charac-terized the rate of decrease of expected distortion with SNR athigh channel SNR range. The distortion exponent, , is definedas the asymptotic decay rate of the average end-to-end distor-tion with increasing channel signal-to-noise ratio (SNR), i.e.,

.In [13], [14] we provide the distortion exponent of a cooper-ative system for a given number of source coding layers andbandwidth ratio, for various transmission strategies such asprogressive, simultaneous and hybrid digital-analog schemes.While [13], [14] only provides distortion exponent results basedon an high SNR analysis, our results here confirm their validityfor a wide range of channel SNR values.

In order to study the effect of finite order modulation, wealso consider BPSK. While it is not possible to obtain an ex-plicit expression for , the capacity of BPSK sig-naling over AWGN channel, it can be easily computed numeri-cally. Note that, when we constrain the transmission to BPSK,the channel capacity is limited by 1 bits per channel use, nomatter how high the channel SNR is. This, in the end, results in anonzero lower bound for the minimum achievable expected dis-tortion (ED) as opposed to the Gaussian codebook case, whereED approaches to zero (or equivalently to infinity) withincreasing channel SNR.

For BPSK modulation we first consider the asymmetric sce-nario with and . Fig. 4compares the reconstructed signal quality (in terms of )of four modes for the asymmetric scenario. We vary(hence and ) to analyze the effect of channelqualities on cooperation. Since the channel capacity is bounded


Fig. 4. Signal SNR, SNR (dB) versus average channel SNR, SNR(dB) for different modes of communication. We assume a bandwidthratio of b = 4 with BPSK modulation, and an asymmetric scenario withSNR = SNR = 4SNR . Results are based on information theoreticanalysis.

above by 1 even when SNR is high, all four modes converge toa similar for large channel SNRs. However, for smallerSNRs, gains due to cooperation are significant. Cooperativecoding (mode 3) improves the end-to-end distortion over modes1 and 2 up to while layered cooperation(mode 4) extends the benefits of cooperative transmission toall SNRs. Comparison with Fig. 3 illustrates that fixing themodulation scheme limits the potential improvements one canget with layered cooperation. Therefore, adaptive modulation(based on average channel SNRs) coupled with layered cooper-ation would be essential for better reconstructed signal quality.

To study the effects of various link qualities, we also con-sider the symmetric case for which in Fig. 5.To understand the effect of inter-user channel between and

, we consider in Fig. 5(a) (perfect inter-user channel) and in Fig. 5(b). We chose

as an intermediate value. Note that, we examinebetween 0–30 dB. Hence, for

and have worse quality channels to thanthe channel, while for , ob-serves a better channel to than to . Note that, perfor-mances of mode 2 and mode 1 are not affected by the inter-userchannel. Comparing the two plots in Fig. 5, we observe that bothmode 3 and mode 4, perform very close to the perfect inter-userchannel case even when the inter-user channel has a fixed av-erage SNR of 14 dB. Note that in Fig. 5(b) cooperative coding(mode 3) performs better than direct transmission (mode 1) be-yond , that is even when channel hasworse quality than channel. This is due to the increaseddiversity of cooperative transmission [4]. Also, while mode 3and 4 reconstructed signal qualities are slightly worse than theasymmetric case, we still get significant improvements over nocooperation (modes 1 and 2). Outage or error probability of co-operation is usually smaller for asymmetric scenarios such asthe one studied here, mainly due to improvements in pathloss,which is consistent with our distortion analysis.

Fig. 5. Signal SNR, SNR (dB) versus average channel SNR (dB) for dif-ferent modes of communication based on information theoretic analysis. Weassume a bandwidth ratio of b = 4 with BPSK modulation, and a symmetricscenario with SNR = SNR . (a) Perfect inter-user channel and (b) 14 dBinter-user channel.

D. Simulation Results for Practical Channel Codes

In order to study the performance achievable by practicalchannel coding and modulation schemes, we use BPSK mod-ulation and RCPC convolutional codes [22]. A rate 1/4 mothercode with (17, 13, 13, 15) generator is employed. The code canthen be punctured to rates 1, 2/3, 1/2, and 1/3. At the receiver, aViterbi decoder is used for decoding. We fix our channel framesize, to be 1728 bits and the number of samples per frame,

to be 432, resulting in . Sincethis results in a bandwidth ratio of . The optimal bit al-location problem explained in Section II is solved through ex-haustive search, that is, we obtain the various probabilities in(2)–(7) through channel simulations for every possible bit al-location satisfying the rate constraint and find the optimal bitallocation that minimizes ED.


Fig. 6. Signal SNR,SNR (dB) versus average channel SNR (dB) of practicalchannel coding scheme (RCPC) for both the asymmetric and the symmetricscenarios, and b = 4. The symmetric scenario assumes the inter user channelis perfect.

Similar to the information theoretic analysis, we examine thedecoded signal SNR both for an asymmetric and a symmetriccooperation scenario ( and

with perfect inter-user channel, respec-tively). The results for both scenarios are plotted in Fig. 6,where the reconstructed signal quality (in terms of SNRg)obtained with optimal bit allocation using the four communica-tions modes is plotted as a function of increasing . Thetrends are similar to the information theoretic bounds in Fig. 4and 5, and is slightly lower than the bounds as expected.By examining the optimal bit allocation resulting from ourstudy, we find that, at low , the need for more protectionthrough channel coding leads to no bit assignment for EL. Somode 4 operates as mode 3 (and mode 2 as 1), which explainsthe overlap of mode 4 and mode 3 (mode 2 and 1) at the low

region in Fig. 4. At high , channel coding bitsare no longer necessary, hence the optimal allocation assignsall bits to source compression to reduce compression distortion.Therefore, mode 4 and mode 2 (mode 3 and 1) overlap at high

in Fig. 4. Layered transmission modes are equivalent,or better than single layer modes (mode 2 is better than 1, andmode 4 is better than 3). We have also carried out simulationsfor , and found out that the optimal bitallocations are identical to perfect inter-user channel case.

Our results for the BPSK modulation scheme (both informa-tion theoretic and the practical coding results) show that i) lay-ered compression and transmission (mode 2) brings improve-ment over direct transmission (mode 1) at high range;ii) Cooperative channel coding (mode 3) improves performanceover mode 1 and mode 2 up to medium values of , butmode 3 is not as good as mode 2 at high ; iii) layeredcooperation (mode 4), by combining the benefits of mode 2 andmode 3, provides the best performance over all other modesfor the entire range of tested; and iv) lastly, there is adistinct improvement from mode 4 over all other modes in amid-range of .

IV. COOPERATION FOR VIDEO TRANSMISSION

A. Video Source and Encoder

In the previous section, we showed the fundamental bene-fits of layered cooperation for i.i.d Gaussian sources, and ar-gued that end-to-end expected distortion obtained with prac-tical codes closely resembles the information theoretic boundsfor fixed modulation. Motivated by this, in this section we per-form experiments with real video signals and practical channelcodes. Note that although the distortion-rate function of a videoencoder generally still follows the exponential decay behavior[26], [27], the impact of packet losses on the end-to-end dis-tortion is far more complicated. This is because errors injectedat different locations of a video stream contribute differently tothe final distortion. Hence the expected distortion expression forGaussian sources in (1) will not hold for video signals. Althoughthere has been prior research (e.g. [28], [29]) on modeling videodistortion from channel packet loss, these models are developedunder various assumptions about the video signal characteris-tics and the channel error statistics. Instead of using analyticalmodels, we choose to code real video sequences and simulatetheir transmissions.

For video encoding, we use an H.263+ codec [24] with SNRscalability. The H.263+ encoder uses the popular block-basedmotion compensated prediction (MC). Each macroblock (MB)in the current video frame is expressed in terms of the bestmatching MB in the past frame through a motion vector (MV),which describes the block translation. In place of the color inten-sities, only the MVs and the prediction error are transmitted. MCimproves coding efficiency by exploiting dependency betweenadjacent video frames, but it makes the compressed stream verysensitive to channel errors. The use of previously reconstructedvideo frames at the decoder forms a prediction loop and intro-duces temporal error propagation. One way to circumvent thisproblem is to use periodic intra refresh, where an MB is codedperiodically without MC using an intra mode. This allows theMB to be decoded independently. Error propagation thus stopsat the next received intra MB.

In our experiment, we generate the compressed video bitstreams using the H.263+ encoder software [30] and simulateits transmission over wireless medium for all of the four com-munication modes. We intra-code the first video frame in thetest sequence, called I-frame. All the other video frames, calledP-frames, are compressed using MC from one previously re-constructed video frame. For each P-frame, we intra-code everyJ-th MB, while all other MBs are coded in an inter mode usingMC. We define the intra rate as . Note that hereafter,we use just “frame” to refer to a “video frame” while we willuse “channel frame” to specifically refer to a channel frame.

Layered or scalable compression is achieved by partitioningvideo data into different layers. BL contains the essential in-formation, while ELs contain additional information for refine-ment. Different ways of partitioning the video data constitutesthe different scalability modes in H.263+: temporal, SNR andspatial. For the simulations of mode 2 and mode 4, we chooseSNR scalability. The BL is a low but acceptable quality repre-sentation of video obtained by using a relatively large quanti-zation step size, and the EL is the difference between the orig-


inal frame and the BL, represented with a smaller quantizationstep size. In our setup, besides the first I-frame, the BL is de-pendent on one previous BL frame; while the EL is dependenton the current BL frame as well as one previous EL frame.Thus a loss in the EL bitstream will not affect the reconstruc-tion of the BL frames. Scalability is an error resilience featuresince the more essential BL can be prioritized when networkor system resources are limited. Other error resilience features[31], [32] offered in H.263+ include the use of resynchroniza-tion markers. In our experiment, we group the MBs into group ofblock (GOB), with each GOB containing one row of MBs. Weadd a resynchronization marker at the beginning of each GOB.The GOB header is useful when decoding a video frame thatspans multiple channel frames. If a channel frame is lost, onlythe affected GOBs are lost (replaced by “0” bits). The video de-coder will search for the next GOB header and resume decodingfrom there. Hence, instead of losing the entire video frame, someGOBs can still be decoded. For those MBs that cannot be de-coded, we use a simple frame copy error concealment methodthat copies the corresponding block from the previously recon-structed frame.

We fill the channel frames with encoded video bits accordingto the channel frame bit allocation assuming negligible lengthpacket header at the physical layer. We put successivecompressed BL bits and successive compressed EL bitsinto their respective layers. Generally there is no alignmentbetween GOBs and channel frames. The channel coder thenapplies channel coding according to the modes described inSection II. Given a fixed channel frame rate frames/s, for acandidate source bit allocation of and bits per channelframe, the video bit rates are bits/s for BL and

bits/s for EL. We use the rate control algorithmin [30] to achieve the desired bit rates. However, the resultingbit rate is not exactly the same as the target rate. Although wechoose the simulation parameters so that the target number ofbits per video frame (including source and channel coding bits)will fit an integer number of channel frames, the number ofsource bits produced per video frame generally varies. As wepacketize the compressed video into channel frames, we start anew video frame in a new channel frame, which implies that thelast channel frame in a video frame may not be fully packed.Stuffing the unused bits with “0”s, we actually operate at a bitrate slightly higher than that specified by the bit allocation.Another assumption we make is that the first video frame in thesequence, as well as all video frame headers are uncorruptedduring transmission. In practice, this may be achieved by aguaranteed out-of-band channel. Instead of minimizing ex-pected distortion as we did for the Gaussian source, we quantifyvideo quality with the popular objective video quality measure,average peak signal to noise ratio (PSNR) of the decoded video.The framewise PSNR is defined as

where MSE is the mean squared error between the original anddecoded luminance component of a video frame. We averagethe PSNR over all transmitted and decoded frames. Therefore italso includes the averaging over channel fading. We maximize

the decoded video PSNR jointly over source coding ( , and), channel coding ( , ), and cooperation .

B. Video Simulation and Results

For the FER simulations, we use the same channel code, mod-ulation level, and channel frame length as in Section III-D. Forthe Gaussian source, the coherence time of the channel is not im-portant as long as it is a multiple of channel frames. For video,however, coherence time determines how many video framesobserve the same fading level, thus affecting the quality of thereconstructed signal. We assume a coherence time of 0.1 s. For3G, 802.11 WLAN, and WIMAX systems, each with carrier fre-quencies in the GHz range, this would model a pedestrian user.We fix our channel frame rate F to be 100 frames/s and our videoframe rate to be 10 video frames/s (fps). Therefore each videoframe is sent with approximately 10 channel frames, duringwhich the fading level stays constant. Because each channelframe carries 1728 bits, the target total video bit rate (includingsource and channel coding) is 172.8 kbits/s (Kbps). This bit rateis appropriate for low resolution video transmission over wire-less channels (where a significant portion of the bit rate may beused for channel coding).

We carry out experiments on three standard test video se-quences in QCIF format (176 144 Y pixels per frame). Thesequences are: A low-motion sequence “Mother & Daughter”(300 frames); an intermediate motion sequence “Foreman” (300frames, with large panning at the end of the sequence); and ahigh-motion sequence “Football” (208 frames). Skipping 2 outof every 3 frames, the original 30 fps test sequences are con-verted to 10 fps and compressed for transmission. We loop eachvideo sequence as necessary to transmit 10 000 channel frames.Considering total bit rate of 172.8 Kbps, we vary the BL bit ratefrom 43.2 to 172.8 Kbps, and the EL rate from 0 to 86.4 Kbps.

Our results for “Foreman” using symmetric cooperationare shown in Fig. 7 for perfect inter-user

channel and for . The trends are similarto the Gaussian source trends. Layered cooperation (mode 4)consistently outperforms direct transmission (mode 1), andimproves over other modes at different average ranges.When the inter-user channel is perfect, the improvement ofcooperation over direct transmission is more pronounced forvideo than for Gaussian source. Better channel robustnessfrom cooperation reduces not only distortion in the currentlyaffected video frame, but also that in the future frames whichare dependent upon this frame through MC. Note that noisein the inter-user channel dampens the benefits of cooperationmore for video than for i.i.d. Gaussian source where the dif-ference between and wasinsignificant. The reason behind this still needs to be thoroughlyexplored. One possible explanation is that the inefficiency ofthe practical source encoder limits the strength available forerror protection in the inter-user channel. Another reason is thatany residual channel error lowers the decoded video qualitymore significantly than for the Gaussian case, because of theerror propagation effect.

For video delivery, we are mainly interested in improvingthe video quality at low to medium range because at


Fig. 7. Average end to end quality (in terms of PSNR) of “Foreman” videostream versus average channel SNR, SNR in dB with BPSK modulation andRCPC channel coding for different communication modes. We have symmetriccooperation with SNR = SNR . (a) Perfect inter-user channel and (b) 14dB inter-user channel.

higher , the quality is already acceptable without coop-eration. As shown in Fig. 7, user cooperation (mode 3 and mode4) brings significant improvement to video quality (in terms ofPSNR) at low to medium range. Furthermore, mode 4(layered cooperation) offers distinct improvement over all othermodes in the intermediate range.

We show the corresponding optimal bit allocations in Table Ifor mode 4 at perfect . As expected, as in-creases, less error protection is needed and hence generally bothchannel coding bits (including cooperation bits) and the intrarate decreases. However, when a “sharp transition” occurs in thebit allocation with a significant drop in the channel coding bits(e.g. from to and fromto in Table I), increases slightly to offer moreprotection against error propagation. We can conclude that theparameter here acts to “fine tune” the quality. We note that thisdiscontinuity in the optimal values for the coding parameters

TABLE IOPTIMAL BIT ALLOCATION AND � FOR VIDEO “FOREMAN” AT MODE 4,

SYMMETRIC COOPERATION, PERFECT INTER-USER CHANNEL

may be due to the fact that we did not search over a continuousrange for each parameter, which is computationally prohibitive.

We carry out the same experiments on two other video se-quences: the lower-motion sequence “Mother & Daughter” andthe higher-motion sequence “Football”. Fig. 8 shows the resultsfor the symmetric scenario with perfect inter-user channel case.Besides the expected higher PSNR for the low-motion “Mother& Daughter” (and vice versa for the faster “Football”), we maketwo observations on the trend as increases from low tohigh: First, the layered modes (mode 2 and 4) start to strictlydominate the single-layer modes (mode 1 and 3) at a lower

when the video has higher motion. Higher motion videorequires more source bits but at the same time, it is also moreaffected by channel errors because distortion from error-prop-agation is more severe. Layering therefore offers the best so-lution by assigning more bits for the video source through EL,while maintaining the channel coding rate to protect the essen-tial BL. We conclude from here the importance of layered com-pression with UEP for high motion video. Our second obser-vation is that, the cooperation modes (3 and 4) converge to theno-cooperation modes (1 and 2) at a lower for highermotion video. Since the final video distortion is contributed byboth channel errors and lossy video compression, as the amountof motion increases, the compression distortion is more severefor the same source coding rate. Compared to low motion video,the channel-error induced distortion thus becomes less domi-nant for higher motion sequence at a lower channel .

In the case of asymmetric cooperationfor video, we measure video PSNR for the “Foreman”

sequence as varies from 0 to 30 dB. The results areplotted in Fig. 9. The trends are similar to the Gaussian sourceand the end-to-end PSNR is higher than the symmetric case.

It is of interest to find out if the optimal bit allocation pro-vides substantial PSNR gain over one that is not optimized. Forboth mode 3 and mode 4, we choose the optimal bit alloca-tion and for “Foreman” for , a mid-range

. Then, in Fig. 10, we plot the PSNR for the entire rangeof using this parameter set for symmetric cooperationwith perfect inter-user channel. Also present in the graph is the


Fig. 8. Average end to end quality (in terms of PSNR) of “Football”, and“Mother & Daughter” video streams as a function of channel SNR, SNR(dB), with BPSK modulation and RCPC channel coding for different commu-nication modes. We have symmetric cooperation (SNR = SNR ) withperfect inter-user channel.

Fig. 9. Video PSNR for Foreman, QCIF, 10 fps, versus channel SNR, SNR(dB) for both asymmetric and symmetric scenarios. The symmetric scenarioassumes perfect inter-user channel.

PSNR result of optimizing only , but not the bit allocation. Theresults show that optimization provides substantial gains (up to6 dB) in PSNR for both mode 3 and mode 4 at low .When is high, the gain is less, but still significant (upto 3 dB). Therefore adaptation of the source and channel codingand cooperation based on channel conditions can bring signif-icant gains. Furthermore, the system performance is less sen-sitive to the mismatch between the actual channel SNRs andassumed average channel SNRs in the high SNR range, but ismore sensitive to the mismatch in the low SNR range.

V. CONCLUSIONS

User cooperation is a popular spatial diversity technique thathas been previously explored from a physical layer perspective.

Fig. 10. Optimization versus no optimization comparison for the quality (interms of PSNR) of Foreman video sequence. “Fixed” corresponds to the resultusing parameters optimized for SNR = 12 dB in a symmetric, perfect inter-user channel setup, (a) compares the results for single layer cooperative coding,(b) compares results for layered cooperation. (a) Mode 3 and (b) mode 4.

In this paper, we explore a cross layer approach and utilize usercooperation to maximize the end-to-end source quality. We opti-mize the parameters of source coding, channel coding and coop-eration jointly to minimize the expected source distortion at thereceiver. Making use of the increased channel reliability offeredby cooperation, we introduce a layered cooperation scheme thatprotects the more important base layer through cooperation andstronger channel coding.

Our information theoretic analysis for an i.i.d. Gaussiansource and Gaussian channel codebooks (no constraint on max-imum modulation level) suggests that cooperative coding andlayered cooperation are superior to noncooperative strategiesfor a wide range of bandwidth ratios and average link SNRs. Infact layered cooperation outperforms all other modes in termsof end-to-end average distortion and provides a higher rate ofdecrease in the distortion as a function of channel SNR. Theresults in this paper are consistent with [13], [14], where we


concluded that in communication over fading channels, layeredsource coding dramatically improves the minimum expecteddistortion for high SNRs. When the modulation is constrained,we observe that for both i.i.d. Gaussian source and videosource, single layer cooperation can significantly improve thesource quality for low to medium channel SNR values. Athigher channel SNRs, layered compression with UEP throughchannel coding is necessary for further improvements. Bycombining user cooperation and layered compression, we areable to improve the source quality for a larger channel SNRrange. The layered cooperation scheme also provides a distinctimprovement over either cooperation alone or layering alonefor a mid range of channel SNRs. Our results are confirmed byboth information theoretic analysis based on outage probabilityof the transmission and simulation of a practical modula-tion/channel coding scheme.

The importance of layering is particularly profound for highmotion video. Our study illustrates that for symmetric users,when a good quality inter-user cooperative channel is present,cooperation provides a slightly better improvement for a videosource when compared with a Gaussian source. However, thepresence of temporal error propagation in the video decoder alsocauses the performance to be more sensitive to the inter-userchannel. On the other hand, if the relay terminal can provide abetter quality channel to the user’s destination, as in our asym-metric cooperation, video quality can be enhanced greatly by co-operation or layered cooperation even if the inter-user channelis not perfect.

Finally, we would like to comment on how adaptation ofbit allocation among source coder, channel coder and cooper-ation can be accomplished in practical systems. Note that ouroptimization problem is formulated based on average receivedchannel SNR’s rather than instantaneous fading levels. There-fore, we can precompute the optimal bit allocations for differentcombinations of , , and , and store theresults in a look up table at the sender. We assume that the valuesof , , and can be estimated at the re-ceivers (relay and the destination) and fed back to the senderperiodically, and the sender will choose its bit allocation ac-cording to the look-up table. We further assume the sender caninform the relay and the destination its bit allocation using someauxiliary data packets. Note that, the transmission of the SNRand bit allocation information requires a very low rate link sinceaverage channel SNR’s change at a very low rate. Because theoptimal bit allocations do not need to be computed in real time,their computational complexity is not a limiting factor in thesystem design.

REFERENCES

[1] M. Etoh and T. Yoshimura, “Advances in wireless video delivery,”Proc. IEEE, vol. 93, no. 1, pp. 111–122, Jan. 2005.

[2] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diversity-part I: System description,” IEEE Trans. Commun., vol. 51, no. 11, pp.1927–1938, Nov. 2003.

[3] A. Sendonaris, E. Erkip, and B. Aazhang, “User cooperation diver-sity-part II: Implementation aspects and performance analysis,” IEEETrans. Commun., vol. 51, no. 11, pp. 1939–1948, Nov. 2003.

[4] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative diversityin wireless networks: Efficient protocols and outage behavior,” IEEETrans. Inform. Theory, vol. 50, no. 12, pp. 3062–3080, Dec. 2004.

[5] T. Hunter and A. Nosratinia, “Diversity through coded cooperation,”IEEE Trans. Wireless Commun., vol. 5, no. 2, pp. 283–289, Feb. 2006.

[6] T. Hunter, S. Sanayei, and A. Nosratinia, “Outage analysis of codedcooperation,” IEEE Trans. Inform. Theory, vol. 52, no. 2, pp. 375–391,Feb. 2006.

[7] A. Stefanov and E. Erkip, “Cooperative coding for wireless networks,”IEEE Trans. Commun., vol. 52, no. 9, pp. 1470–1476, Sep. 2004.

[8] J. N. Laneman and G. W. Wornell, “Distributed space-time coded pro-tocols for exploiting cooperative diversity in wireless networks,” IEEETrans. Inform. Theory, vol. 49, no. 10, pp. 2415–2525, Oct. 2003.

[9] A. Wittneben, “Coherent multiuser relaying with partial relay cooper-ation,” in Proc. IEEE Wireless Communication and Networking Conf.,Las Vegas, NV, Apr. 2006.

[10] O. Shalvi, “Multiple source cooperation diversity,” IEEE Commun.Lett., vol. 8, no. 12, pp. 712–714, Dec. 2004.

[11] D. Gündüz and E. Erkip, “Joint source-channel cooperation: Diversityversus spectral efficiency,” in Proc. of 2004 ISIT, Chicago, IL, Jun.2004.

[12] X. Xu, D. Gündüz, E. Erkip, and Y. Wang, “Layered cooperative sourceand channel coding,” in Proc. IEEE Int. Conf. Communications, Seoul,Korea, May 2005.

[13] D. Gündüz and E. Erkip, “Source and channel coding for cooperativerelaying,” in Proc. 2005 Int. Workshop on Signal Processing Advancesfor Wireless Communications (SPAWC), New York, Jun. 2005.

[14] D. Gündüz and E. Erkip, “Source and channel coding for cooperativerelaying,” IEEE Trans. Inform. Theory, to be published.

[15] X. Xu, Y. Wang, and E. Erkip, “Layered cooperation for wireless mul-timedia communications,” in Proc. 2004 Picture Coding Symp., SanFrancisco, CA, Dec. 2004.

[16] H. Shutoy, Y. Wang, and E. Erkip, “Cooperative source and channelcoding for wirless video transmission,” in Proc. IEEE Int. Conf. ImageProcessing (ICIP), Atlanta, GA, Oct. 2006.

[17] J. N. Laneman, E. Martinian, G. W. Wornell, and J. G. Apostolopoulos,“Source-channel diversity for parallel channels,” IEEE Trans. Inform.Theory, vol. 51, no. 10, pp. 3518–3539, Oct. 2005.

[18] A. Kwasinski, Z. Han, and K. J. R. Liu, “Joint source-channel codingfor cooperative multimedia communications,” in Proc. 2005 Int. Work-shop on Signal Processing Advances for Wireless Communications(SPAWC), New York, Jun. 2005.

[19] W. Equitz and T. Cover, “Successive refinement of information,” IEEETrans. Inform. Theory, vol. 37, no. 2, pp. 269–275, Mar. 1991.

[20] L. H. Ozarow, S. Shamai, and A. Wyner, “Information theoretic con-siderations for cellular mobile radio,” IEEE Trans. Veh. Technol., vol.43, pp. 359–378, May 1994.

[21] L. Zheng and D. Tse, “Diversity and multiplexing: A fundamentaltradeoff in multiple antenna channels,” IEEE Trans. Inform. Theory,vol. 49, pp. 1073–1096, May 2003.

[22] J. Hagenauer, “Rate compatible punctured convolutional codes (RCPCcodes) and their cpplications,” IEEE Trans. Commun., vol. 36, no. 4,pp. 389–400, Apr. 1998.

[23] ITU-T. Recommend. H.263, Video Coding for Low Bit Rate Commu-nication 1998.

[24] G. Cote, B. Erol, M. Gallant, and F. Kossentini, “H.263+: Video codingat low bit rates,” IEEE Trans. Circuits, Syst., Video Technol., vol. 8, no.7, pp. 849–866, Nov. 1998.

[25] T. M. Cover and J. A. Thomas, Elements of Information Theory. NewYork: Wiley, 1991.

[26] L. Choi, M. T. Ivrla, E. Steinbach, and J. A. Nossek, “Sequence-levelmodels for distortion-rate behavior of compressed video,” in Proc.IEEE. Conf. Image Process., Sep. 2005, vol. II, pp. 486–489.

[27] Z. He and D. Wu, “Resource allocation and performance analysis ofwireless video sensors,” IEEE Trans. Circuits, Syst., Video Technol.,vol. 16, no. 5, pp. 590–599, May 2006.

[28] K. Stuhlmüller, N. Färber, M. Link, and B. Girod, “Analysis of videotransmission over lossy channels,” IEEE J. Select. Areas Commun., vol.18, no. 6, pp. 1012–1032, Jun. 2000.

[29] Y. Wang, Z. Wu, and J. M. Royce, “Modeling of transmission-loss-induced distortion in decoded video,” IEEE Trans. Circuits, Syst., VideoTechnol., vol. 16, no. 6, pp. 716–732, Jun. 2006.

[30] Signal Processing and Multimedia Lab., Univ. British Columbia, TMN8 (H.263+) Encoder/Decoder, Version 3.1.3 TMN 8 (H.263+) codec,Feb. 1998.

[31] Y. Wang, S. Wenger, J. Wen, and A. K. Katsaggelos, “Error resilientvideo coding techniques,” IEEE Signal Processing Mag., vol. 17, no.4, pp. 61–82, Jul. 2000.

[32] S. Wenger, G. Knorr, J. Ott, and F. Kossentini, “Error resilience supportin H.263+,” IEEE Trans. Circuits, Syst., Video Technol., vol. 8, no. 6,pp. 867–877, Nov. 1998.


Hoi Yin Shutoy received the B.S. degree in 1999from Polytechnic University, Brooklyn, NY, and theM.S. degree in 2004 from Columbia University, bothin electrical engineering. She was a Ph.D. candidateat Polytechnic University at the time of the paper.

Deniz Gündüz (S’02) received the B.S. degree inelectrical and electronics engineering from MiddleEast Technical University, Ankara, Turkey, in 2002and the M.S. degree in electrical and computer engi-neering from Polytechnic University, Brooklyn, NY,in 2004. He is currently pursuing the Ph.D. degree atPolytechnic University.

In 2004, he was a Summer Researcher with theLaboratory of Information Theory (LTHI), EPFL,Lausanne, Switzerland. His research interests lie inthe areas of communications theory and information

theory with special emphasis on joint source-channel coding and cross-layerdesign.

Elza Erkip (S’93–M’96–SM’05) received the theB.S. degree in electrical and electronics engineeringfrom Middle East Technical University, Ankara,Turkey, and the M.S. and Ph.D. and degrees inelectrical engineering from Stanford University,Stanford, CA.

She joined Polytechnic University, Brooklyn, NY,in Spring 2000, where she is currently an AssociateProfessor of electrical and computer engineering.From 1996 to 1999 she was with the Departmentof Electrical and Computer Engineering, Rice Uni-

versity, Houston, TX. Her research interests are in wireless communications,information theory, and communication theory.

Dr. Erkip received the 2004 Communications Society Stephen O. Rice PaperPrize in the Field of Communications Theory, the NSF CAREER award in2001 and the IBM Faculty Partnership Award in 2000. She is an Associate Ed-itor of IEEE TRANSACTIONS ON COMMUNICATIONS, a Publications Editor ofIEEE TRANSACTIONS ON INFORMATION THEORY, and a Guest Editor of IEEESignal Processing Magazine, Special Issue on Signal Processing for Multiter-minal Communication Systems. She is the Technical Area Chair for the “MIMOCommunications and Signal Processing” track of 41st Annual Asilomar Confer-ence on Signals, Systems, and Computers, and the Technical Program Co-Chairof 2006 Communication Theory Workshop.

Yao Wang (M’90-SM’98-F04) received the B.S.and M.S. degrees in electronic engineering fromTsinghua University, Beijing, China, in 1983 and1985, respectively, and the Ph.D. degree in electricaland computer engineering from the University ofCalifornia at Santa Barbara in 1990.

Since 1990, she has been with the faculty of Poly-technic University, Brooklyn, NY, and is presentlyProfessor of electrical and computer engineering.She was on sabbatical leave at Princeton University,Princeton, NJ, in 1998 and at Thomson Corpo-

rate Research, Princeton, in 2004–2005. She was a Consultant with AT&TLabs-Research, formerly AT&T Bell Laboratories, from 1992 to 2000. Herresearch areas include video communications, multimedia signal processing,and medical imaging. She is the lead author of the textbook Video Processingand Communications and has published over 150 papers in journals andconference proceedings.

Dr. Wang has served as an Associate Editor for IEEE TRANSACTIONS ON

MULTIMEDIA and IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO

TECHNOLOGY. She received New York City Mayor’s Award for Excellence inScience and Technology in the Young Investigator Category in year 2000. Shewas elected Fellow of the IEEE in 2004 for contributions to video processingand communications. She is a co-winner of the IEEE Communications SocietyLeonard G. Abraham Prize Paper Award in the Field of Communications Sys-tems in 2004.

ieee journal of selected topics in signal...

Documents