Error Resilience and Recovery in Streaming of. Embedded Video

Transcription

1 Error Resilience an Recovery in Streaming of Embee Vieo Sungae Cho an William A. Pearlman a a Center for Next Generation Vieo Research Rensselaer Polytechnic Institute 110 Eighth Street Troy, NY Abstract The three-imensional (3-D) SPIHT coer has prove its efficiency an its realtime capability in compression of vieo. A forwar-error-correcting (FEC) channel (RCPC) coe combine with a single ARQ (automatic-repeat-request) prove to be an effective means for protecting the bitstream. There were two problems with this scheme: the noiseless reverse channel ARQ may not be feasible in practice; an, in the absence of channel coing an ARQ, the ecoe sequence was hopelessly corrupte even for relatively clean channels. In this paper, we first introuce a newpartitioning metho of wavelet coefficients to achieve error resilience. All of the spatio-temporal (s-t) tree blocks correspon to the whole image sequences, because we get wavelet coefficients in some fixe intervals over the whole wavelet coefficients. This proceure brings higher error resilience, since we can conceal the Preprint submitte to Elsevier Science 26 September 2001

2 lost coefficients with the surrouning coefficients even if some of substreams are totally missing. The bitstreams are packetize an encoe with a channel coe. Then, we present an unequal error protection of vieo bitstreams using three imensional SPIHT (3-D SPIHT) algorithm an our propose 3-D SPIHT algorithm. We emonstrate that the 3-D SPIHT compresse bitstreams can also be error resilient against channel bit errors by unequally protecting the embee vieo bitstreams, an more error resilient when we protect the substreams of our propose algorithm unequally. Key wors: SPIHT, Error Resilient Vieo Coing, Unequal Error Protection, Vieo Compression, Embee Bitstream 1 INTRODUCTION Wavelet zerotree image coing techniques were evelope by Shapiro (EZW) [2], an further evelope by Sai an Pearlman (SPIHT)[3], an have provie unpreceente high performance in image compression with low complexity. Later, Kim et al. extene the iea to the three imensional SPIHT (3-D SPIHT)algorithm [4,5] that employe a temporal wavelet transform instea of motion compensation, an compare the result with the vieo compression algorithms which are generally use nowaays, such as MPEG-2 an H.263. They showe promise of a very effective an computationally simple vieo coing, an also obtaine excellent results numerically an visually. However, wavelet zerotree coing algorithms are, like all algorithms proucing variable length coewors, extremely sensitive to bit errors. A single-bit transmission error may lea to loss of synchronization between encoer an ecoer execution paths, which woul lea to a total collapse of ecoe vieo 2

3 quality. To achieve robust vieo over noisy channels, the work of Sherwoo an Zeger [8,9] was extene to progressive vieo coing by Kim et al. They have shown in Ref. [14,15] that the robustness is increase when we cascae the 3-D SPIHT coer with rate-compatible puncture convolutional (RCPC)channel coer [1] an automatic repeat request (ARQ)to protect the 3-D SPIHT bitstream from being corrupte by channel errors. This approach increases robustness, but is still susceptible to early ecoing failure. Another approach towar protecting vieo ata from channel bit errors is to moify the 3-D SPIHT algorithm to work inepenently in a number of socalle spatio-temporal (s-t)blocks that are ivie into fixe-length packets an interleave to eliver a fielity embee output bit stream. This algorithm is calle STTP-SPIHT (Spatio-Temporal Tree Preserving 3-D SPIHT) [6]. As a result, any bit error in the bitstream belonging to any one block oes not affect any other block, so that higher error resilience against channel bit errors is achieve. Therefore any early ecoing failure affects the full extent of the GOF in the normal 3-D SPIHT, but in the STTP-SPIHT, the failure allows reconstruction of the associate region with lower resolution only. This algorithm gives excellent result in most cases, but may still experience very early ecoing errors, resulting in lower resolution vieo in specific regions. This metho was inspire by Ref. [11 13] in the fiel of image coing area with EZW algorithm [2]. In Ref. [7], Alatan et al. showe the embee image bitstreams can be elivere with error resilience maintaine by iviing the bitstreams into three classes. They protect the subclasses with ifferent channel coing rates of 3

4 the RCPC coer [1], an improve the overall performance against channel bit errors. In this paper, we use the iea in [6], but a ifferent metho of partitioning the wavelet coefficients into s-t blocks to get higher error resilience. Instea of grouping ajacent coefficients, we group coefficients at some fixe interval in the lowest subban, an this interval is etermine by the number of s-t blocks S. Then we track the spatio-temporal relate trees of the coefficients, an merge them together. As a result, the s-t blocks of the STTP-SPIHT correspon to certain regions, but the s-t blocks of our new metho correspon to the whole region with lower resolution. We call this algorithm Error Resilient an Error Concealment 3-D SPIHT (ERC-SPIHT)algorithm. Then, we show how the 3-D SPIHT encoe vieo bitstreams can be implemente to unequal error protection by sub-iviing the embee bitstreams, presenting a hybri coer which combines the ERC-SPIHT algorithm an unequal error protection. This metho can effectively protect the very early ecoing error, because we protect more strongly the beginning portion of the bitstream. The organization of this paper is as follows: Section 2 shows error resilient 3-D SPIHT algorithm. Section 3 shows unequal error protection of the 3-D SPIHT algorithm. Section 4 provies simulation results. Section 5 conclues this paper. 4

5 a a b c c c c a b b 1 3 c c c c c c c c Fig. 1. Structure of the spatio-temporal relation of 3-D SPIHT compression algorithm 2 Error Resilient 3-D SPIHT Algorithm 2.1 System Overview Figure 1 shows how coefficients in a three-imensional (3-D)transform are relate accoring to their spatial an temporal omains. Character a represents a root block of pixels (2 2 2), an characters b, c, enote its successive offspring progressing through the ifferent spatial scales an numbers 1, 2, 3 label members of the same spatio-temporal tree linking successive generations of escenants. We use 16 frames in a GOF (group of frames), therefore we have 16 ifferent frames of wavelet coefficients. We can observe that these frames not only have spatial similarity insie each one of them across the ifferent scales, but also temporal similarity between frames, which will be efficiently exploite by the Error Resilient an Error Concealment SPIHT (ERC-SPIHT)algorithm. 5

6 a b a b aabbaaaabbbb c c aabbaaaabbbb a b a b ccaaaabbbb c c ccaaaabbbb aabbaabbcccc aabbaabbcccc cccccccc cccccccc aaaabbbbaaaabbbb aaaabbb bb aaaabbbb aaaabbb a a a a bbbb aaaabbbbaaaabbbb cccc cccc cccccccc cccccccc cccccccc aa aaaaaa aa aaaaaa aa aaaaaa aa aaaaaa aaaa aaaa aaaa aaaa aaaa aaaa aaaa aaaa b b b b bb bbbb bb bb bb bb bbbb bbbb bbbbbb bbbb bbbb bbbbbbbb bbbbbbb b bbbb bbbb cc cccccc cc cccccc cc cccccc cc cccccc cccccccc cccccccc cccccccc cccccccc 3D SPIHT 3D SPIHT 3D SPIHT 3D SPIHT Encoer Encoer Encoer Encoer Stream 1 Stream 2 Stream 3 Stream 4 Block Interleaver STTP-SPIHT bit stream Fig. 2. Structure of the Spatio-Temporal Tree Preserving 3-D SPIHT(STTP-SPIHT) compression algorithm In Ref. [6], we reporte the Spatio-Temporal Tree Preserving 3-D SPIHT(STTP- SPIHT)compression algorithm. Figure 2 shows the structure an the basic iea of the STTP-SPIHT compression algorithm. The STTP-SPIHT algorithm ivies the 3-D wavelet coefficients into some number S of ifferent groups accoring to their spatial an temporal relationships, an then to encoe each group inepenently using the 3-D SPIHT algorithm, so that S inepenent embee 3-D SPIHT substreams are create. These bitstreams are then interleave in blocks. Therefore, the final STTP-SPIHT bitstream will be embee or progressive in fielity, but to a coarser egree than the normal SPIHT bitstream. In this figure, we show an example of separating the 3-D wavelet transform coefficients into four inepenent groups, enote 6

7 by a, b, c,, each one of which retains the spatio-temporal tree structure of normal 3-D SPIHT [4,5], an these trees correspon to the specific regions of the image sequences. The s-t block, which is enote by a, matchesthe top-left portion in all frames of the sequence transform. The other s-t blocks correspon to the top-right, bottom-left, bottom-right fractions of the image sequences, an those s-t blocks are enote by b, c,, respectively. The normal 3-D SPIHT algorithm is just a case of S = 1, an we can flexibly choose S. When we choose the number of substream S for the image size of X Y,each of X an Y shoul be ivisible by 2 L+1,whereL is the number of ecomposition levels. If the axis is not ivisible by the number, we shoul exten the original image to be ivisible. For the Football sequence with 3 levels of ecomposition, we can choose certain values of S from 1 up to 330. STTP-SPIHT gave us excellent results in both noisy an noiseless channel conitions while preserving all the esirable properties of the 3-D SPIHT [6]. However, this metho is also susceptible to early ecoing error, an this error results in one or more small regions with lower resolution than the surrouning area. Sometimes, this artifact occurs in an important region. To avoi this, early ecoing error shoul be prevente so as to guarantee a minimum quality of the whole region. We can use a ifferent metho for partitioning the wavelet coefficients into s-t blocks to solve the problem. The 3-D SPIHT compression kernel is inepenently applie to each tree composing the wavelet coefficients in the lowest subban an the spatially relate coefficients in the higher frequency subbans. The algorithm prouces sign, location an refinement information for the trees in each pass. Therefore, we nee to keep the spatio-temporal relate trees to maintain compression efficiency. However, we o not have to 7

8 Conventional Grouping New Metho of Grouping Fig. 3. Graphical illustration of the 2 methos (STTP-SPIHT an ERC-SPIHT) keep together the contiguous wavelet coefficients in the lowest subban, since the kernel is inepenently applie to each tree roote in a single lowest subban coefficients an branching into the higher frequency subbans at the same s-t orientation. In our propose algorithm, therefore, we group the lowest subban coefficients at some fixe interval instea of grouping ajacent coefficients. This interval is etermine by the number of s-t blocks S, the image imensions, an number of ecomposition levels. Then, we track the spatio-temporal relate trees of the coefficients, an merge them together. As a result, the s-t blocks of the STTP-SPIHT correspon to certain regions, but the s-t blocks of our new metho correspon to the whole region with lower resolution. We call this algorithm Error Resilient an Error Concealment STTP-SPIHT (ERC-SPIHT)algorithm. Figure 3 graphically compares the two methos. One nice feature of the ERC-SPIHT is that the very early ecoing failure affects the whole region because the ecoe coefficients woul be sprea out to the whole area along with the sequence, an the coefficients are conceale by the other surrouning coefficients which are ecoe at a higher rate. When the ecoing failure occurs in the same position, the quality of ERC-SPIHT is much better than that of STTP-SPIHT in visually an numerically (PSNR)because ERC-SPIHT algorithm itself has the function of 8

9 error concealment. Therefore, the ERC-SPIHT no longer suffers from small areas which are ecoe with a very low resolution. Figure 4 shows the worst case example of this iea. We use Football sequence, an coe at 1.0 bit/pixel with ERC-SPIHT (P = 16). We assume the ecoing error occurre in the beginning of the substream number 2 (secon packet), so that one of the substreams is totally missing. In this figure, (a), (c) are the result from the ERC-SPIHT without error concealment, an (b), () are the result with error concealment. In this case, we just use the average values of surrouning coefficients for the missing coefficients only in the root ban. As we can see, in the case of without error concealment, there are many black spots in the images. However, when we use concealment scheme for the missing coefficients, we can recover the missing area very well. 2.2 Error Resilience The most important characteristic of SPIHT coer is embeeness in fielity, which means that we can get higher image quality with more uncorrupte bits. Therefore, we measure the error resilience in this section by the number of clean packets receive. Our metho hols the similar characteristic of error resilienceas in [8,9,11 13]. We use the Viterbi algorithm to ecoe each packet, an the ecoing error of the Viterbi algorithm occurs inepenently for each packet receive. If we say M 1 as the total number of packets in the normal SPIHT, then, the each of the ERC-SPIHT substream has M S (= M 1 )packets. S We can express the probability of ecoing error occurring in one of the first 9

10 Fig. 4. coe to 1.0 bit/pixel, with ERC-SPIHT (P = 16). (a)top-left : Football sequence, without concealment (frame 1), PSNR = (b)top-right : Football sequence, with concealment (frame 1), PSNR = B, (c)secon row-left : without concealment (frame 16), PSNR = B, ()Secon row-right : with concealment (frame 16), PSNR = B M S packets, p 1 (M S ), of the bitstream in terms of ξ. Hence, we have p 1 (M S )=1 Prob{No FailureinM S packets} =1 (1 ξ) M S (1) M S ξ, where ξ is the probability of ecoing error of Viterbi algorithm for each packet. From this, we can see that the longer the bit-stream, the more likely it is to be ecoe incorrectly, in other wors, if M S is small, then the p 1 (M S ) 10

11 will be also small. Hence, we have p 1 (M S ) p 1 (M 1 ), for S > 1 (2) We can also fin the expecte number of uncorrupte packets using ξ(1 ξ) k, 0 k<m S p 2 (k) = (3) (1 ξ) k, k = M S where p 2 (k)is the probability of successful ecoing of first k packets of a substream. Then, we can fin the average number of uncorrupte packets using m s = M S k=0 k p 2 (k)(4) where m s is an average number of uncorrupte packets for substream number s. Therefore, total number of uncorrupte packets receive is S m s.when the channel conition is favorable, S m s = m 1. When channel conition is highly unfavorable, m s m 1, then, there are approximately S times more clean packets which are correctly ecoe. Therefore, we can express as S m s m 1, for S > 1 (5) From the equations above, we know that partitioning wavelet coefficients to inepenent substreams gives more clean packets than that of original bitstream. 11

12 Average PSNR Bit Error Position x 10 4 Fig. 5. Bit error sensitivity of the 3-D SPIHT compresse bitstream of monochrome Football sequences (first 16 frames) coe 1.0 bpp (bit per pixel) 3 UNEQUAL ERROR PROTECTION OF EMBEDDED BITSTREAMS 3.1 Unequal Error Protection Of The 3-D SPIHT The 3-D SPIHT compression kernel is compose of 2 passes, a sorting pass an a refinement pass, repeately performe until the total bits prouce meet the bit buget. From the sorting pass, sign bits an location bits are prouce, an from the refinement pass, refinement bits are generate. The location bits are results of significance tests on sets of pixels, incluing singleton sets, an are often calle the significance map. As Alatan et al i in Ref. [7], we classify the bits into two classes accoring to their bit error sensitivities. The sign bits an refinement bits can be classifie as sign an refinement bits (SRB), an the location bits can be classifie by themselves (LOB). If any bit error occurs in LOB, then the compresse bit 12

13 Average PSNR Bit Rate Fig. 6. Average PSNRs versus bitrates for monochrome Football sequences stream is useless after the point where the bit error occurs. However, any bits in the SRB which are affecte by channel bit errors o not propagate as long as the LOBs are error free. From our experimental results, the size of the SRB ranges from 20% to 25% of the original bitstream, epening on the rate. Figure 5 shows bit error sensitivity of compresse bitstream encoe by the 3-D SPIHT algorithm. In this figure, we compresse the first sixteen frames of monochrome Football sequences, an put one bit error to the compresse bitstream from beginning to the en of the bitstream. In this figure, the x-axis represents the bit error position, an the y-axis means the average PSNR of ecoe sequence. As we can see, only one bit error in the beginning of the bitstream affects the whole bitstream, an makes the bitstream meaningless. In aition to that, we can see many positions where the average PSNRs are very high compare to the other positions. This calss is calle SRB (Sign an Refinement Bit), since these bits represent sign an 13

14 No Image Sequences Wavelet Transform Sorting Pass Refinement Pass Meet Byte Buget? Yes Stop Refinement Location Bits Bits Sign Bits SRB(Sign & Refinement Bits) Without Arithmetic Coing LOB(Location Bits) Arithmetic Coing Final Bitstream Fig. 7. Unequal error protection of the 3-D SPIHT algorithm refinement information of wavelet coefficients. The other class is also calle LOB (Location Bits), because these bits contain the information of location of the wavelet coefficients an shoul be synchronize between encoer an ecoer. In aition, the 3-D SPIHT algorithm has an important property that all the compresse bits are positione in the orer of importance. This means that SPIHT prouces a purely embee or progressive bitstream, meaning that the later bits in the bitstream refine earlier bits, an the earlier bits are neee for the later bits to be useful. Figure 6 represents average PSNR values versus bitrates for monochrome Football sequence. From this figure, we can see that the average PSNR value very rapily increases when bitrates are lower than 0.05 bpp. Above this rate, PSNR increases at much more graually with bitrate. This result implies that the very beginning of the bitstream shoul be more strongly protecte against channel bit errors, than later portions of the bitstream. 14

15 SRB LOB1 LOB rate 1 rate 2 rate 3 Final Bitstream r LOB1 < rlob2 < r SRB Fig. 8. Bit rate assignment of the unequal error protection For this reason, even if we have only the beginning part of the bitstream, we can still get a rough renition of the source. But if we lose just a small portion at the beginning part of the bitstream containing LOB bits, then we can not reconstruct anything from the bitstream. From this iea, we can further subpartition the LOB class to two classes LOB1 an LOB2, corresponing to the earlier an later parts, respectively, of the bitstream. We can utilize these facts to get higher error resilience against channel bit errors. We separate the SRB an LOB in the original bitstream, an transmit the SRB first with lowest error protection (highest channel coe rate), then LOB1 an LOB2, each with stronger protection (lower channel coe rate) than SRB, but with LOB1 receiving a lower coer rate (higher protection) than LOB2. Figure 7 an Figure 8 graphically illustrate this iea. Figure 7 shows the structure of the unequal error protection 3-D SPIHT (UEP-SPIHT) an how the bits are classifie an combine together. As we can see, we o not use the arithmetic coing for SRB bits to avoi error propagation among the bits. Figure 8 presents the bitrate assignments accoring to their bit error sensitivities an importance. As we can see in this figure, LOB1 shoul be highly protecte, because these bits are more important than others in terms of bit sensitivities an the orer of importance, then LOB2 an SRB can be protecte with successively higher channel coing rates. 15

16 SRB 1 SRB 2 SRB 3 SRB n LOB 1 LOB 2 LOB LOB n r1 SRB r2 r LOB Fig. 9. Bitstream of STTP-SPIHT algorithm As epicte in the Figure 8, we sen the SRB bits first, then sen LOB bits next. This means that while sening SRB bits, this bitstream is not progressive. However, after sening SRB bits, this bitstream is purely progressive, since all the SRB bits are store in buffer, an these bits are use with LOB bits together. As we mentione before, the size of SRB bits ranges from 20% to 25% of the total bitstream for source coe rates about 1 bpp (2.53 Mbps). Therefore we get higher error resilience against channel bit errors while we sacrifice the progressiveness to a small extent. In the UEP-SPIHT heaer, we nee just one negligible aitional item of information, the SRB size. 3.2 Unequal Error Protection With The ERC-SPIHT ERC-SPIHT gave us excellent results in both noisy an noiseless channel conitions while preserving all the esirable properties of the 3-D SPIHT. However, this metho still stops ecoing process for the substream where the first ocoing error occurs. If the ecoing error occurs at very early stage, we can conceal the ata. In aition to that, this early eoing error coul be effectively protecte by unequal error protection metho. To implement unequal error protection to ERC-SPIHT, we partition the sub- 16

17 bitstreams accoring to their bit sensitivities an the orer of importance. Figure 9 shows us this iea. Each substreams are ivie into SRB an LOB, enote by SRB1 - SRBn, anlob1-lobn. Asweuseinthe3-DSPIHT algorithm,wesenallthesrbsfirst,anthensenlobs.unlikeinthecase of 3-D SPIHT, we use block interleaving/einterleaving scheme for LOB area to maintain progressiveness. The overhea of this metho is the information bits which are save in each sub-bitstream heaer to convey its SRB size. The ecoer reas the heaer first, an put the SRBs to buffer areas accoring to the information of the SRBs size as the bits are arriving. Once all the SRBs have arrive, the ecoer einterleaves the LOBs accoring to the packet size, since the LOBs are sent as interleave bitstream, an ecoes the bitstreams with SRBs together. Therefore, early portions of this bitstream are protecte strongly with little loss of progressiveness. 3.3 Source/Channel Coing Rate Figure 10 shows the system escription of 3D/ERC-SPIHT with RCPC coer. The functions of block interleaving an einterleaving are neee only for the ERC-SPIHT. Before RCPC encoing, we partition the bitstream into equal length segments of N bits. Each segment of N bits is then passe through a cyclic reunancy coe (CRC)[16,17] parity checker to generate c parity bits. In a CRC, binary sequences are associate with polynomials of a certain polynomial g(x)calle the generator polynomial. Hence, the generator polynomial etermines the error control properties of a CRC. Next, m bits, where m is the memory size of the convolutional coer, are 17

18 3D/ERC- Block SPIHT Interleaving CRC+RCPC Noisy Channel Viterbi Decoer + CRC Check Deinterleaving ERC- SPIHT return channel(optional) Fig D/ERC-SPIHT with RCPC system framework pae at the en of each N + c + m bits of the segment an passe through the rate r RCPC channel coer, which is a type of puncture convolutional coer with the ae feature of rate compatibility. TheeffectivesourcecoingrateR ef f for the original 3-D SPIHT is given by R ef f = Nr N + c + m R total, (6) where a unit of R ef f an R total can be either bits/pixel, bits/sec, or the length of bit-stream in bits. The total number of packets M is calculate by R ef f /N, where R ef f is the bitstream length. In the case of unequal error protection, we first get the R ef fsrb an M SRB accoring to the Equation(6). Then R ef flob1 an R ef flob2 can be calculate by r LOB1 N + c + m R LOB1 + r LOB2 N + c + m R LOB2 = M M SRB, (7) where R LOB1 + R LOB2 = R total - R SRB. Therefore, we can get R LOB1 an R LOB2. Table 1 shows the relative percentage of source coing bits an channel coing bits using this equation for monochrome Football sequences at total transmission rate of 2.53 Mbps when BER (Bit Error Rate) is

19 r Source Channel 3-D SPIHT 2/3 60% 40% SRB 4/5 14% 5.5% UEP- LOB1 4/7 14% 13.5% SPIHT LOB2 2/3 32% 21% Table 1 Relative percentages of source/channel (RCPC) bits at total transmission rate of 2.53 Mbps when BER (Bit Error Rate) is RESULT In our test of error resilience, we assume that the channel is binary symmetric (BSC). For MPEG-2, we use 15 frames in a GOP (Group Of Pictures), an the I/P frame istance is 3 (IBBPBBP...). For the 3D/ERC-SPIHT, sixteen frames in a GOF (Group Of Frames)are use, an a yaic three level transform using 9/7 biorthogonal wavelet filters [20] is applie to the image sequences. For the 3-D SPIHT an MPEG-2, we sen the bitstream sequentially in 200 bit packets, an for the ERC-SPIHT, we interleave the streams in 200 bit packets to maintain embeeness, an the receiver e-interleaves the bitstream to a series of substreams, each one of which is ecoe inepenently. The algorithm is then teste using the monochrome Football sequences. The istortion is measure by the peak signal to noise ratio (PSNR) PSNR =10log 10 ( MSE ) B, (8) 19

20 Fig Football sequence (frame15) (a) Top-left : Original sequence (b) Top-right : using ERC-SPIHT (n=16)/rcpc with BER = PSNR = B (c) Bottom-left : ussing MPEG-2/RCPC with BER = PSNR = B () Bottom-right : using ERC-SPIHT with unequal error protection with BER = PSNR = B where MSE enotes the mean square error between the original an reconstructe image sequences. All PSNR s reporte for noisy channels are averages over twenty (20)inepenent runs. As in previous works [8,9,14,15], we protecte the 200 bit packets with the CRC, c = 16 bit parity check while generator polynomial g(x) =X 16 + X 14 + X 12 +X 11 +X 8 +X 5 +X 4 +X 2 +1, an RCPC channel coer with constraint length m = 6. We focuse on bit error rates (BER)of ɛ = 0.01 an 0.001, because the BER s of most wireless communication channels are ɛ = We set the total transmission rate R total to 2.53 Mbps, r = 2/3 for ɛ = 20

21 0.01 an 8/9 for ɛ = for the equal error protection (EEP). In our case of unequal error protection (UEP), we use a certain r only for LOB2, an use the one level higher rate available to us for the SRB, an one level lower rate for the LOB1 with the same transmission rate of 2.53 Mbps. In the case of ɛ = 0.01, the RCPC rate for LOB2 = 2/3, an the rate for the SRB = 4/5, an the rate for the LOB1 = 4/7. Once we ecie the channel coing rate, we can easily calculate the bit buget for the three classes using Equation (6) an (7). At the estination, the Viterbi ecoing algorithm [18,19] is use to convert the packets of the receive bit-stream into a 3D/ERC-SPIHT an MPEG-2 bit-stream. In the Viterbi algorithm, the best path chosen is the one with the lowest path metric that also satisfies the checksum equations. In other wors, each caniate trellis path is first checke by computing a c =16 bit CRC. When the check bits inicate an error in the block, the ecoer usually fixes it by fining the path with the next lowest metric. However, if the ecoer fails to ecoe the receive packet within a certain epth of the trellis, it stops ecoing for that 3D/ERC-SPIHT stream. For MPEG-2, which is not embee an nees the full bitstream to see the whole frames, when ecoing failure occurs, we can use one of two schemes. One is just to use the corrupte packet, an the other is to put all 0 s to the corrupte packet. In this paper, we use the corrupte packet itself. Figure 11 shows the comparison of Football sequence with RCPC an BER = In this figure, (a)is the original Football sequence (frame 15). Typical reconstructions of Football sequence at total transmission rate of 2.53 Mbps an channel bit error rates of 0.01 with S =16an MPEG-2 are shown in Figure 11 (b), (c) an (). As we can see, in Figure 11, 21

22 (b), the ERC-SPIHT stops ecoing for the substream where ecoing failure occurs. Therefore any early ecoing failure affects to the whole region, an the coefficients are conceale by the other surrouning coefficients which are ecoe at higher rates. In this example, the very early ecoing error occurs at 10th substream, but it is very har to iscern the affecte region. However, the MPEG-2 ecoe sequence in Figure 11 (c), the ecoing failure affects some block, an the block is fille with some other picture s block. In the case of unequal error protection ERC-SPIHT, the probability of very early ecoing error is very low because this metho strongly protects the earlier part of the bitstream, an usually gives a nice result. In this figure 11, (c)is the typical example of ecoe sequence (frame 15). Table 2 shows the comparison of average PSNRs among 3-D SPIHT, UEP- SPIHT, an UEP/ERC-SPIHT for monochrome Football sequence at total transmission rate of 2.53 Mbps. As we can see, the average PSNR of UEP-SPIHT is about 2-5 B higher than those of the equal error protection 3-D SPIHT (EEP-SPIHT). In the hybri metho, when BER is 0.01 the average PSNR is B, which is a little higher than that of equal error protection of ERC-SPIHT, an when BER is 0.01 the average PSNR is B. In aition to the higher PSNR, we have prevente very early ecoing error using the lower channel coing rate for the LOB1. The merit over the unequal error protection of the 3-D SPIHT is that we can get consistent results. In other wors, the unequal error protecte 3-D SPIHT still stops ecoing whenever ecoing failure occurs, however, the unequal error protecte ERC- SPIHT prevents very early ecoing error effectively, an still operates until S ecoing failures occur. 22

23 BER EEP 3D-SPIHT/RCPC 24.5 B 28.2 B EEP ERC-SPIHT/RCPC B B UEP 3D-SPIHT/RCPC B B UEP ERC-SPIHT/RCPC B B MPEG/RCPC B B Table 2 Comparison of average PSNR over 20 inepenent trials among 3-D SPIHT, UEP- SPIHT, an UEP/ERC-SPIHT at total transmission rate of 2.53 Mbps 5 CONCLUSION We have implemente unequal error protection for the 3-D SPIHT an the ERC-SPIHT algorithm. The result is very promising while sacrificing small egree of progressiveness. The most remarkable thing is we effectively mitigate the early ecoing error for the ERC-SPIHT. References [1] J. Hagenauer, Rate-compatible puncture convolutional coes (RCPC coes) an their applications, IEEE Transactions on Communications, vol. 36, pp , April [2] J. M. Shapiro, Embee image coing using zerotrees of wavelet coefficient, IEEE Transactions on Signal Processing, vol. 41, pp , December [3] A. Sai an W. A. Pearlman A New, Fast an Efficient Image Coec Base on 23

24 Set Partitioning in Hierarchical Trees, IEEE Trans. on Circuits an Systems for Vieo Technology, vol. 6, pp , June [4] B.-J. Kim an W. A. Pearlman An embee wavelet vieo coer using threeimensional set partitioning in hierarchical trees, Proc. of Data Compression Conference, pp , March [5] B.-J. Kim, Z. Xiong, an W. A. Pearlman Low Bit-Rate Scalable Vieo Coing with 3D Set Partitioning in Hierarchical Trees (3D SPIHT), IEEE Trans. Circuits an Systems for Vieo Technology, vol. 10, pp , December [6] S.ChoanW.A.PearlmanError Resilient Compression an Transmission of Scalable Vieo, Applications of Digital Image Processing XXIII, Proceeings SPIE, vol. 4115, pp , also IEEE Trans. Circuits an Systems for Vieo Technology (to be publishe). [7] A. Ayin Alatan, M. Zhao, an A. N. Akansu Unequal Error Protection of SPIHT Encoe Image Bit Streams, IEEE Jounral on Selecte Areas In Communications, vol. 18, pp , June [8] P.G.SherwooanK.Zeger Progressive Image Coing for Noisy Channels, IEEE Signal Processing Letters, vol. 4, pp , July [9] P. G. Sherwoo an K. Zeger Progressive Image Coing on Noisy Channels, Proc.DCC, pp , April [10] H. Man, F. Kossentini, an M. J. T. Smith Robust EZW Image Coing for Noisy Channels, IEEE Signal Processing Letters, vol. 4, no. 8, pp , August [11] C. D. Creusere A New Metho of Robust Image Compression Base on the Embee Zerotree Wavelet Algorithm, IEEE Transactions on Image Processing, vol. 6, no. 10, pp , October

25 [12] C. D. Creusere Robust image coing using the embee zerotree wavelet algorithm, Proc. Data Compression Conference, pp. 432, March [13] C. D. Creusere A family of image compression algorithms which are robust to transmission errors, Proc. SPIE, vol. 2668, pp , January [14] Z. Xiong, B.-J. Kim, an W. A. Pearlman Progressive vieo coing for noisy channels, Proc. IEEE International Conference on Image Processing (ICIP 98), vol. 1, pp , October, [15] B.-J. Kim, Z. Xiong, an W. A. Pearlman, an Y. S. Kim Progressive vieo coing for noisy channels, Journal of Visual Communication an Image Representation, vol. 10, pp , [16] T. V. Ramabaran an S. S. Gaitone A Tutorial on CRC Computations, IEEE Micro, vol. 8, pp , August [17] G. Castagnoli, J. Ganz, an P. Graber Optimum Cyclic Reunancy-Check Coes with 16-Bit, IEEE Transactions on Communications, vol. 38, pp , January [18] G.D. Forney, Jr. The Viterbi algorithm, Proc. IEEE, vol. 61, pp , January [19] N. Seshari an C. Sunberg List Viterbi ecoing algorithm with applications, IEEE Transactions on Communications, vol. 42, pp , [20] M. Antonini, M. Barlau, P. Mathieu, an I. Daubechies Image coing using wavelet transform, IEEE Transactions Image Processing, vol. 1, pp ,