Text from the interactive explanation
Turbo-detected unequal protection MPEG-4 wireless video telephony using multi-level coding, trellis coded modulation and space-time trellis coding
S.X. Ng, J.Y. Chung and L. Hanzo
Original paperHow close can a live video call get to the limits of radio?
Two phones receive the same video call through the same fading radio channel. One guards every bit equally; the other guards the bits that matter most. Turn down the power and watch which call survives.
Eb/N0 is the energy spent on each bit compared with the background radio noise — this page's “transmit power” knob. It is measured in decibels (dB) , a ratio scale: +3 dB is roughly double the power, +10 dB is ten times. Trellis states count the decoder's work per pass; both phones get the same budget.
Equal protection (benchmark)
Unequal protection (the paper's design)
“1 in N” is the bit error ratio (BER): the fraction of bits that arrive flipped. Picture quality is the paper's PSNR, a fidelity score in dB comparing the received picture with the original — higher is closer, and 39.7 dB here means no transmission errors at all.
Slide the power down until the right-hand call starts to break. Then find where the left call becomes clean.
SourcesBit error ratios and picture quality are re-plotted from Figs 7–12 of Ng, Chung & Hanzo (2005), which are computer simulations over fading radio channels. The caller and the picture damage are an illustrative model built for this page and driven by those error rates; the paper itself shows no corrupted pictures.
A video call is a stream of bits
Every radio link has a hard limit: a minimum energy per bit below which nothing can be received reliably. The question behind this study is how close a live video call, sent over fading airwaves, can get to that limit. The researchers' answer starts with a simple observation: not all bits are equal.
A bit is a single 1 or 0. The video phone in the study squeezes each picture into a small packet of them: a postage-stamp image of 176 × 144 pixels (the “QCIF” format), 30 pictures a second, sent at 69,000 bits per second. That works out at 2,300 bits per picture .
On the way, radio noise and fading — the signal swelling and collapsing as waves bounce around buildings and cancel each other — flip some of those bits. The fraction that arrive flipped is the bit error ratio: a BER of 10−4 means 1 bit in 10,000 is wrong.
Engineers fight back with error-correcting codes. They add redundancy: extra, carefully calculated bits that let the receiver spot and repair flips. A code's code rate is the share of useful bits in what is sent: rate 1/2 means one extra bit for every data bit. Redundancy is not free — every extra bit costs radio bandwidth or transmit power.
So with a fixed budget, the real question is: which bits deserve the most protection?
Not all bits are equal
Inside each picture's 2,300 bits, a few do a very different job from the rest.
The video is cut into packets — here, one per horizontal stripe of the picture, a row of 16 × 16-pixel tiles called macroblocks. Each packet starts with a short header: framing and synchronisation bits that tell the decoder where the packet belongs and how to read it. The rest is picture detail: patterns of brightness and colour, plus motion hints that say which part of the previous picture to reuse.
Predict first. You flip one single bit. Which hurts the picture more: a bit in a packet header, or a bit of picture detail?
There is no wrong answer — make a guess, then test it in the lab below.
Toy modelSimplified error model built for this page; in the opening scene the error rates come from the paper. Damage stays until you send a key frame, or for 4 seconds after your last flip.
Tap any header strip (gold) or detail strip (teal) to flip one bit in that packet.
Why the headers matter so much
Framing and synchronisation bits are only about 10% of the bitstream, but an error there makes the decoder lose its place — and because each picture is built from the one before, the damage lingers until a fresh key frame redraws everything. A flipped detail bit, by contrast, usually spoils one small block.
So the researchers sorted every picture's bits into two classes. Class 1 — 25%, or 575 bits — holds the framing bits plus the most sensitive ~15% of the picture-detail bits. Class 2 — the other 75%, or 1,725 bits — holds the remaining detail and motion bits.
Armour where it matters
If a quarter of the bits matter far more than the rest, give that quarter thicker armour — without sending more in total.
1 · Same budget, spent differently
The paper uses multi-level coding: a separate error-correcting code for each class. Class 1 gets a strong rate-1/2 code (one extra bit for each data bit); class 2 gets a lighter rate-3/4 code (one extra bit for every three). The total length barely changes — the armour is simply moved to where it counts.
The paper tried two families of codes for each class: non-systematic convolutional codes ( NSC ) and recursive systematic convolutional codes ( RSC ). Both are classic “sliding-window” codes that mix each data bit into several transmitted bits.
2 · The payoff
Does the thicker armour work? In the paper's simulations, the important class-1 bits come through with more than an order of magnitude lower bit error ratio than the ordinary class-2 bits at the same transmit power. Drag across the chart, or use the slider. (Iterations are repeated passes of the receiver's decoders over the same picture — Chapter 5 shows how they work.)
Re-plottedRe-plotted from Fig 8 of Ng, Chung & Hanzo (2005); computer simulations over uncorrelated Rayleigh fading channels. Curves stop where the published plot ends at 10−5.
3 · A second layer: set partitioning
After coding, the bits ride on radio symbols. The paper uses 16-QAM: 16 distinct points — a “constellation” — each standing for 4 bits and sent as a particular strength and timing of the wave. Noise nudges the received point; if it lands nearer a neighbour, bits come out wrong.
Trellis coded modulation (TCM) splits the 16 points into families whose members are farther apart at every split : d, then √2·d, 2d and 2√2·d. The code's protected bits pick the family; inside a family the points are so far apart that noise rarely confuses them. The paper maps the class-1 bits onto the subsets with the largest minimum distance — extra protection exactly where it matters.
Press Split and watch the mix-up rate fall for the highlighted family. Raise the noise to make it harder.
Simulation for this pageRandom noise added to the highlighted family's points; a mix-up is counted when a noisy point lands nearer another member of the same family. The dashed line marks the closest pair.
Air that fades
Radio waves reach a phone along many paths at once — bouncing off walls, cars and buildings. Sometimes they add up; sometimes they cancel.
The result is fading: the signal strength swells and collapses, many times a second. In a deep fade the whole constellation shrinks into the noise and even a strong code struggles. The paper's answer is antenna diversity: two transmit and two receive antennas, tied together by a space-time trellis code (STTC). With four separate radio paths, it is unlikely that all of them fade at the same moment.
Simulation for this pageEach radio path is a randomly varying (“Rayleigh”) channel, slowed right down so you can watch it; with several antennas the strength shown is the average over all transmit–receive paths. The paper's simulations assume fading that changes independently from one symbol to the next.
The twist: a steadier channel changes the best design
With enough diversity, the paper notes, the fading envelope becomes nearly constant — “Gaussian-like”. That changes which way of combining coding and modulation works best. Bit-interleaved coded modulation (BICM) was designed for channels that fade from symbol to symbol; TCM was designed for steady channels with only background noise.
Once the space-time code has steadied the link, TCM wins clearly. Switch between them:
Re-plottedRe-plotted from Figs 9 and 10 of Ng, Chung & Hanzo (2005); computer simulations over uncorrelated Rayleigh fading channels. The benchmark's 8-iteration curve is not in the paper, so its best plotted curve (4 iterations) is shown.
Three decoders in conversation
At the receiver, three decoders take turns. Each makes its best guess about the bits, then passes on only what it newly learned — its “extrinsic information” — as hints for the others. Round after round, the guesses sharpen. This is turbo decoding, and each round is one iteration.
The STTC decoder untangles the signals from the two antennas; the TCM decoder reads the constellation families; the MLC decoder runs the two class codes. Certainty is measured in bits of information per symbol: up to 4 for STTC and TCM (each 16-QAM symbol carries 4 bits), up to 3 for the MLC.
SchematicAnimation built for this page, shaped to match the end points the paper reports at Eb/N0 = −0.5 dB with a 10,000-symbol interleaver (Figs 4–5): the STTC decoder's output rises only from 2.0148 to 2.3493 of 4, while three stages converge to full certainty. It is not the paper's computed EXIT surfaces.
Why a third stage?
The paper's analysis found that the STTC decoder on its own can only climb from about 2.01 to 2.35 bits of certainty out of 4 — even when handed perfect hints. In a conventional two-stage receiver, like the benchmark's, that is a dead end at very low power: the conversation stalls with errors left over. With the third stage, the hints have somewhere to grow, and the paper's charts predict full certainty — error-free decoding at Eb/N0 = −0.5 dB .
To check this, engineers draw EXIT charts, which track how much each decoder's output improves as its input hints improve. The paper's novelty was computing them in three dimensions for non-binary symbols, by averaging symbol probabilities — and concluding that with this kind of space-time code, a three-stage receiver is needed to approach the channel's capacity .
An interleaver is a shuffler that spreads neighbouring bits apart before sending, so that a burst of errors arrives as scattered single errors the codes can fix. Longer shufflers decode better but add delay — Chapter 7 returns to this.
A fair fight: counting the work
Decoding work is counted in trellis states per iteration — the number of bookkeeping states each decoder tracks. The paper's three small decoders cost 40 per iteration; the benchmark's single big code costs 80. So two of the paper's iterations cost the same as one of the benchmark's.
The evidence
Everything in the opening scene comes from these curves. Each one is a computer simulation: the authors sent a standard test video, picture by picture, through simulated fading radio channels and counted what came out wrong.
Lower curves are better: they reach a small bit error ratio at less power. Hover or drag across the chart, or use the slider, to read every curve at one power level.
Re-plottedRe-plotted from Figs 9 and 10 of Ng, Chung & Hanzo (2005); computer simulations over uncorrelated Rayleigh fading channels. Bit-error curves stop where the published plots end at 10−5; picture-quality curves start at the bottom of the published plot (15 dB).
How close to the wall?
Back to the opening question. Every radio link has a channel capacity: a theoretical limit, set by the laws of information theory, below which no code at all could deliver the data reliably.
For this two-by-two antenna link sending 1.98 bits per 16-QAM symbol, the paper computes that wall at Eb/N0 = −1.80 dB (below 0 dB means each bit carries less energy than the noise). A simpler two-antenna scheme, the space-time block code (STBC), has a lower ceiling: its wall sits at −0.49 dB.
How close the design gets depends on the interleaver — the shuffler from Chapter 5. A longer shuffle decodes better but delays the picture. Slide from a live call to a broadcast:
Re-plottedCapacity curves re-plotted from Fig 3 of Ng, Chung & Hanzo (2005) (2 × 2 antennas, 16-QAM). Solid: limit when sending 16-QAM symbols; dashed: limit with ideal, unrestricted signals. Operating points from the paper's text (0.5 dB simulated; −0.5 dB and −0.77 dB from its EXIT charts); with the 100,000-symbol shuffler the throughput rises to 2.00 bits per symbol, where the STTC wall is −1.77 dB and the STBC wall −0.45 dB. Delays are approximate conversions (1,160 symbols = one 33 ms video frame). Benchmark: 4 iterations at a BER of 10−4, from Fig 10.
Every decibel counts
For a live call, delay must stay short, so the paper's design sits about 2.3 dB from the wall — yet it still needs about 2.8 dB less power than a strong conventional design doing the same decoding work, roughly half. Where seconds of delay are acceptable, as in broadcasting, it comes within about 1 dB of the limit.
The authors note that even better component codes could be designed with their three-dimensional EXIT charts, at the price of more iterations, more delay and more complexity. The lesson carries beyond 2005-era video phones: when you know which bits matter, you can guard them first — and every decibel saved is transmit power a handset need not spend.
The paper
S.X. Ng, J.Y. Chung and L. Hanzo, “Turbo-detected unequal protection MPEG-4 wireless video telephony using multi-level coding, trellis coded modulation and space-time trellis coding”, IEE Proceedings – Communications, Vol. 152, No. 6, pp. 1116–1124, December 2005. DOI: 10.1049/ip-com:20050236 · PDF of the paper
All results in the paper are computer simulations over uncorrelated Rayleigh fading radio channels.
Charts on this page are re-plotted from data extracted from the paper's figures (Figs 3, 7–12).
The synthetic caller, the picture-damage model, the fading demo, the constellation demo and the decoder animation are illustrations and simulations built for this page. The paper used the standard “Miss America” QCIF test video, which is not reproduced here.
The authors acknowledge financial support from the EPSRC (Swindon, UK) and the EU Phoenix project.
Topics in this iPaper
Engineering
This paper proposes a jointly optimised turbo transceiver design capable of providing unequal error protection for MPEG-4 coding aided wireless video telephony.
Abstract, page 1116
Computer science
the state-of-the-art MPEG-4 video codec was amalgamated with a systematically designed sophisticated turbo transceiver
Section 1, page 1116
unequal error protection
a stronger NSC/RSC is used for protecting the more sensitive video bits, while a weaker NSC/RSC is employed for protecting the less sensitive video bits
Section 1, page 1116
trellis coded modulation
Trellis coded modulation (TCM) [1, 2] constitutes a bandwidth-efficient joint channel coding and modulation scheme
Section 1, page 1116
space time coding
space-time trellis coding (STTC) invoked for the sake of mitigating the effects of fading
Abstract, page 1116
iterative decoding
Extrinsic information was exchanged across three serially concatenated decoder stages
Section 1, page 1116
extrinsic information transfer chart
EXIT charts [22] have been widely used for analysing the convergence behaviour of iterative decoding aided concatenated coding schemes.
Section 4, page 1120
multiple input multiple output
Let us consider a multi-input multi-output (MIMO) system employing Nt transmit antennas
Section 3, page 1118
rayleigh fading
The video performance of the various schemes is evaluated when communicating over uncorrelated Rayleigh fading channels.
Abstract, page 1116
video telephony
turbo transceiver was studied and proposed for MPEG-4 wireless video telephony
Section 6 (Conclusions), page 1123