Multiple Description Image and Video Coding for Noisy Channels

Size: px

Start display at page:

Download "Multiple Description Image and Video Coding for Noisy Channels"

Leo Wade
5 years ago
Views:

1 Multiple Description Image and Video Coding for Noisy Channels Manuela. Pereira Friday 18 of July of h00 Page 1 of 53 Supervisor: Prof. Michel Barlaud and Prof. Marc Antonini

2 1. Multimedia communications Outline Problems and requirements; Classical approaches; Alternative solution - Multiple Descriptions; Channel splitting scheme; MDC: implicit redundancy; New approach: explicit redundancy; MDC advantages MDC problem Page 2 of 53

3 Multimedia communications Problems and requirements High bit rates involved with multimedia; The scarcity of wireless bandwidth; Transmission delay; Requirement: good compression rates. Unreliable channels; The time-varying characteristics of the channel; Page 3 of 53 Requirement: effectiveness in presence of channel failures.

4 Multimedia communications Classical approaches ARQ - Automatic Repeat request (ensure error-free delivery). introduces additional delay. FEC - Forward Error Correction difficulty dealing with burst errors; difficulty adapting to time-varying channel. Page 4 of 53

5 Multimedia communications Alternative solution - Multiple Descriptions Page 5 of 53 MD problem was posed by Gersho, Witsenhausen, Wolf, Wyner, Ziv and Ozarow in 1979 Suppose we wish to send a description of a stochastic process to a destination through a communication network. Assume that there is a chance that the description will be lost. Therefore we send two descriptions and hope that one of them will get through. Each description should be individually good. However, if both get through, then we wish the combined descriptive information to be as large as possible. First theoretical results Witsenhausen, Ozarow, El Gamal and Cover and Wolf, Wyner and Ziv

6 Ö Ç }²T«^. \²T³. À ž P ±.«P Ám Â ' ª P ªÈŸ ž P PÁ ÁÊÉ } p Ÿ ^» ËY P TÈi ³.ÌÍ µ,â8 «6 ^ T «^³n P i«^ìb «B ¾.««²8 ÅÂT ª 8 ± ^. T²\³. À ž Èi ª ŸµT P³ Å ž P Â8 i Áf P³ž o i x ƒ i ÂTÁf p» Î D?EdRXCKM8NPOQRFTSÏdRX'X`hSHS WiM`d v F\[ dfg hm8f6 ƒœfz œ4d?ef[ƒdmð ThQfeÑ] dfg X`h EtWÅKM8NPOQRFTS Multimedia drx è E&Wie }NoN[ dfg[ƒdfsdr[ƒh&wiqy[f\x8 TM`dfKè communications drnpgx ShX`e O FJ TQRN X8FJe8N è EF KM8N \F\X8Xpa½O^](sdfM`è hf Channel Nil è EFTdfM }NNo[ƒgF\X XparWig[ÒgF\ \F\X8X WiM`dfQf]ShX8eOZFHEdf EQf][FTK FTg[FTgoepœ±D?E^hXpaBW lže8ftmyè splitting scheme EF M8F\ \FTKè drnpg Nil'è EF M8X`eÅ[F\X8 TM`dfKè drnpgjä è EF±X8F\ \NPg[Ó[F\X8 TM`dfKè drnpgóũ dfqfq4 \NPg^è M`dfOhe8F Qfdfè è QRF F Ôƒè M W dfgƒlžnpm`s Wiè drnpgjœ u gšè EFÅNPè EFTME&Wig[care UxN dfg[ftkzftg[ftg^e [F\X8 TM`dfKè drnpgxshx`eozf l WiM½WiK&WiM`eCWig[±è E^hXC pwiggnpe4dfg }FTgFTM WiQjOZF,dfg[ƒdfsdR[ƒh&WiQfQf] }NN[cœ I,F\ \No[FTMz l`z} Ö Ø Bg \N[FTM l Ö Ø Õ}NPdfg^e'I'F\ \N[FTM ˆX0 I,F\ \No[FTM' ˆX1 ˆX2 Page 6 of 53 ÙÃdf hm8f ƒœfz Ú:D?EFY E&WiggFTQrX`KQfdfè è dfg KM8NPOQRFTS Trade-off: single description quality - joint descriptions quality *,+n* ÛÓÜ67B2 0 7Ã< Ý8Þr=½5 0Z7B;&ß65n<>; D?EF M8X`eè EF\NPM8FTè dr pwiq?m8f\x`hqfe8xhwikkzfpwimdfg#zp{ à átwig[òè M`]e8NJ 8E&WiM WP Te8FTM`d v F è EF X8FTeNil WP 8EdRFTs WiOQRFâohdfg^è hkqrf\x œãd?eft]#ü FTM8FtKM8NPKZN X8F\[äO^]#b#dfe8X8FTgƒå (R 1, R 2, D 1, D 2, D 0 ) z z

7 Multimedia communications MDC: implicit redundancy Page 7 of 53 For speech Jayant and Christensen For image Multiple Description Scalar Quantization ( Vaishampayan). Multiple Description Transform Coding. Square-Transform Based ( Wang, Orchard, and Reibman). Frame Based ( Goyal, Kovacevic, and Vetterli). These approaches are limited in their ability to adapt to changing transmission conditions.

8 Page 8 of 53 Multimedia communications New approach: explicit redundancy - Ortega et al. 1999, 2000 Level of redundancy can be selected by determining: the number of times a given sample (or wavelet coefficient) is transmitted; how many bits should be used for each of the redundant representations; Advantages using bit allocation: the encoder can adjust itself in a simple manner; the decoder do not require any significant changes to its structure; Provides a very simple mechanisms for adaptation to changing network conditions.

9 Multimedia communications MDC advantages MDC is robust due to the redundancy of the MD of the same source; MDC is scalable as each correctly received description improves the decoder performance; MDC does not require prioritized transmission, as each description is independently decodable; The new MDC approach provides a very simple mechanisms for adaptation to changing network conditions. Page 9 of 53

10 Multimedia communications MDC problem The MDC where created to solve a problem related with packet losses. A challenging task is how to design the MDC coder that can automatically adapt the amount of added redundancy according to underlying channel error characteristics. Y.Wang and Q. Zhu It is this challenge we consider in the present work. Page 10 of 53

11 2. Outline General scheme; Proposed bit allocation for MDC; Proposed model based algorithm; Complete Coding Scheme; Results - Central PSNR vs. side PSNR for still image; Conclusions. Page 11 of 53

12 General scheme ( R T, D M ) Side Decoder ( R 1, D 1 ) Wavelet coefficients MDC Quality or Rate Control bitstream bitstream noisy channel Central Decoder ( R 0, D 0 ) rn Side Decoder ( R 2, D 2 ) Page 12 of 53

13 MDC scheme r N MDC Scalar Quantization Entropy Coder ( R 1, D 1 ) Wavelet MD Bit {q i,1 } coefficients Allocation {q i,2 } Scalar Entropy Quantization Coder ( R 2, ) D 2 Page 13 of 53 (, ) R t D m

14 Proposed bit allocation for MDC Problem statement: Two channels MDC, and balanced MDC. Page 14 of 53 (P ) min D 0 ({q i,1, q i,2 }) Constraints R 1 R T 2 and R 2 R T 2 Penalty D 1 D M and D 2 D M Resolution using Lagrange operators

15 Side rate and distortion modeling J ({q i,1, q i,2 }) = D λ j f(r j ) + j=1 2 µ j g(d j ). j=1 For a source with generalized Gaussian distribution, Parisot 2003 shows N R j = a i R i,j ( q i,j ), for all j {1, 2} i=1 Page 15 of 53 D j = with q i = q i σ i. N w i σi,jd 2 i,j ( q i,j ), for all j {1, 2} i=1

16 Central distortion modeling J ({q i,1, q i,2 }) = D λ j f(r j ) + j=1 2 µ j g(d j ). j=1 D 0 = N w i σi,0d 2 i,0 ( q i,1, q i,2 ). i=1 We propose Page 16 of 53 D i,0 ( q i,1, q i,2 ) = 1 1 [ ( min σ 2 σi,0 2 r N + 1 i,1 D i,1 ( q i,1 ), σi,2d 2 i,2 ( q i,2 ) ) +r N max ( σi,1d 2 i,1 ( q i,1 ), σi,2d 2 i,2 ( q i,2 ) )]

17 õö Illustration LL Resolution m=3 LH Resolution m=2 LH Resolution m=3 HL Resolution m=3 HH Description 1 Resolution m=2 HL Resolution m=2 HH Resolution m=1 HL LL Resolution m=3 LH Resolution m=3 HL Resolution m=3 HH Resolution m=2 Resolution m=2 HL Resolution m=2 Resolution m=1 HL Resolution m=1 LH Resolution m=1 HH LH HH Page 17 of 53 Resolution m=1 LH Resolution m=1 HH LL Resolution m=3 LH Resolution m=3 HL Resolution m=3 HH Resolution m=2 Description 2 Resolution m=2 HL Resolution m=2 Resolution m=1 HL LH HH Resolution m=1 Resolution m=1 Finely coded subband LH HH Coarsely coded subband /wx tkjmonõ ù k 3ÿikyzo9s6{wxüAwznwzsap^s6q å o í ü o=yzo=qnmtkvkvf 3p {knuvqo=qtopo=pzikjwxt 3jþ ntkv vf 3p {kn p o=yxþæšpsa{kop{ 3p {ÆjmoP{tkp { 3prq.ntkvkvF 3p {kn ]šps 3jmnmo=yxþÆšPsA{koP{ wxp+q å o qtís {ko

18 Proposed criterion J ({q i,1, q i,2 }) = + + N w i σi,0d 2 i,0 ( q i,1, q i,2 ) i=1 ( 2 N ) λ j a i R i,j ( q i,j ) R T /2 j=1 2 µ j P j j=1 i=1 Side distortion penalty Page 18 of 53 [ ] 2 Dj D M + (D j D M ) P j =, for all j {1, 2} 2

19 Solution of the problem First order conditions J({q i,1,q i,2 }) q i,1 = 0, J({q i,1,q i,2 }) q i,2 = 0, J({q i,1,q i,2 }) λ 1 = 0, J({q i,1,q i,2 }) λ 2 = 0. Page 19 of 53

20 System a 2 (N + 1) equations and 2 (N + 1) unknowns D i,1 R i,1 ( q i,1 ) = D i,2 R i,2 ( q i,2 ) = w i σ 2 i,1 w i σ 2 i,2 λ 1 a i ( Ci,1 1+r N +µ 1 E 1 ), λ 2 a i ( Ci,2 1+r N +µ 2 E 2 ), N i=1 a ir i,1 ( q i,1 ) R T /2 = 0, N i=1 a ir i,2 ( q i,2 ) R T /2 = 0. Page 20 of 53 1, if min(σi,1d 2 i,1 ( q i,1 ), σi,2d 2 i,2 ( q i,1 )) = σi,jd 2 i,j ( q i,j ) C i,j = r N, otherwise. 2 (D j D M ), if D j > D M E j = 0 otherwise

21 C parameter - optimal solution Compute the MDBA for all possible combinations; Choose the combination that results in the minimal central distortion. N = 10: Nb iterations = 1024; N = 40: Nb iterations = Page 21 of 53

22 Proposed algorithm for C parameter Page 22 of 53 Define S as the set of all possible subbands of description 1 and 2. Perform the following steps: 1. Initialization C i,j = 1, for i = 1..N and j = 1, 2; 2. q i,j, for i = 1..N and j = 1, 2; 3. D i,j ( q i,j ), for i = 1..N and j = 1, 2; 4. search the subband k in S with the highest distortion. 5. set the correspondent C k,j value to r N. 6. redefine the set S as S \ k. 7. if S is not empty, go to step 2.

23 Comparison: Optimal solution vs. Proposed algorithm Optimal solution N = 10: Nb iterations = 1024; P SNR = db, for Lena Image at 1 bpp; N = 40: Nb iterations = Page 23 of 53 Proposed algorithm N = 10: Nb iterations = 10; P SNR = db, for Lena Image at 1 bpp; N = 40: Nb iterations = 40.

24 E F Proposed model based algorithm &(')*+ {C i,j, i = 1..N, j = 1, 2} λ 1, λ 2 µ 1, µ 2, η 0 r N,.-0/ )*+21 R i,j *4356'798 $ : #%;=<?>@< '4A D R B *'4C+5 - ' &(3 f(r j ) 0 I - LM13,.-0/ )*+21 λ new j D E%F A 5GC2H /,.-0/ )*+21 BKJ2-0/ < '4A q i,j R i,j D i,j ( q i,j ) < '4A B!J2-0/ D j C i,j R B *'4C+5 - ' D i,j ( q i,j ) Page 24 of 53 I - C new i,j N C old i,j O LM13,.-0/ )*+21 d j = D j D M µ new j = µ old j + η k d j &(3 g(d j ) 0O I - LM13 P *+)*+ {q i,j, i = 1..N, j = 1, 2} QR567S* J 1 $ : #4TVUXW - <YW[Z 56+]\ W6W - C < +5 - '_^ J2- C1A * J 1

25 Complete Coding Scheme r N Scalar Quantization Entropy Coder, ) D 1 ( R 1 Side Decoder ( R 1, D 1 ) Image 2D DWT MD Bit Allocation ( R T, D M ) {q i,1 } {q i,2 } Scalar Quantization Entropy Coder ( R 2, D 2 ) channel 1 N Quality Control O or channel 2 I Rate Control S E Central Decoder Side Decoder ( R 0, D 0 ) ( R 2, D 2 ) Page 25 of 53

26 Results - Central PSNR vs. side PSNR for still image Lena image compressed to 1 bpp 41,0 40,5 40,0 39,5 Central PSNR 39,0 38,5 Page 26 of 53 38,0 37,5 37,0 Miguel & Riskin Jiang & Ortega Servetto & al Proposed without Ci optimization. Proposed with Ci optimization. 36, Side PSNR

27 Conclusions Page 27 of 53 We proposed a MD scheme based on the Discrete Wavelet Transform (DWT) and an efficient bit allocation technique. The different MD are defined when setting the bit allocation of each subband. We name it Multiple Description Bit Allocation (MDBA). The proposed method with C i,j optimization provides the best results when compared with the best MDC techniques reported to date; When noiseless channel the proposed MDC approximates a single description coding (SDC), without channel coding.

28 3. Outline Complete Coding Scheme; Redundancy parameter estimation; Estimation of H y (x); Channel models and associated redundancies; MDC for Video; Conclusions. Page 28 of 53

29 Complete Coding Scheme rn r N BER Scalar Quantization Estimation Entropy Coder, ) D 1 ( R 1 ( R T, D M ) Side Decoder ( R 1, D 1 ) Video Sequence 2D / 3D Scan Based DWT MD Bit Allocation {q i,1 } {q i,2 } Scalar Quantization (, ) R t D m Entropy Coder ( R 2, D 2 ) channel 1 N Quality Control O or channel 2 I Rate Control S E Central Decoder Side Decoder ( R 0, D 0 ) ( R 2, D 2 ) Page 29 of 53

30 Redundancy parameter estimation We propose r N = H y (x) max p(x) (H(x)). r N = 0 when the channel is noiseless. r N = 1 when a very noisy channel is expected. How to estimate H y (x)? Page 30 of 53 * The equivocation H y (x) is the amount of redundancy that the decoder needs to correct the received message. Shannon 1948.

31 Estimation of H y (x) Proposition 1 min p(x) (H y (x)) max p(x) (H(x)) C max p(x) (H y (x)) p(x) distribution of the input channel symbols; H(x) entropy of the input; C channel capacity: C = max p(x) (H(x) H y (x)); 0 C max p(x) (H(x)). Page 31 of 53 We propose to use r N = max p(x)(h(x)) C. max p(x) (H(x))

32 Page 32 of 53 Channel models and associated redundancies Binary symmetric channel r N = plog 2 p + (1 p)log 2 (1 p) C = 1 + plog 2 p + (1 p)log 2 (1 p), bits/symbol. Additive white Gaussian noise channel r N = 1 B log 2(1 + S N ), for a QPSK modulation 2 C B = log 2(1+γ), bits/symbol; B is the channel bandwidth in symbol/s; γ = S N, where S is the received signal power and N is the AWGN power within the channel bandwidth. Rayleigh channel r N = 1 B log 2e.e N S ( e + ln S N + N S ), for a QPSK modulation 2 C B log 2e.e 1 γ ( e + lnγ + 1 ), bits/symbol (Lee s). γ

33 4. Coder characteristics Add a 1D DWT in the time direction to the 2D DWT. This is named a 3D DWT. Without motion compensation. The 3D scan-based DWT transform allows us to adapt the redundancy parameter to time varying states. Page 33 of 53

34 Gaussian channel simulations ber QCIF silent color video compressed to 200 kbits/s (30 frames/s). Y U V BER BER BER Proposed Method SDC + TC Mean PSNR results Page 34 of 53 The proposed MDC provides a gain of 3 db over a standard method using SDC+TC.

$9@A BDCFEHGI5KJML N9OQPSRUT-V!WXV!<ZY.6X[(=-VT-V!\^](<=- =-[a`vb;c;c0dfe(69`_qgh_jika;cml<)nh\m=-_mgh_moq? pq<)nhr_\^6x 6XV!rsV Y4=+<utFnh: 6vnhrwTxynhr(r=+W7nh`Dc?zc;c @Fe%=+<? {D_69r(8^`x=~}.$

35 Gaussian channel simulations ber! "$#%'&()+*,$# *-.&(,'0/1&(- 243!"$#% &$')( *+#% $#$,-.*,/#01 QCIF silent color video compressed to 200 kbits/s (30 frames/s) Page 35 of ;:(<='>?9@A BDCFEHGI5KJML N9OQPSRUT-V!WXV!<ZY.6X[(=-VT-V!\^](<=- =-[a`vb;c;c0dfe(69`_qgh_jika;cml<)nh\m=-_mgh_moq? pq<)nhr_\^6x 6XV!rsV Y4=+<utFnh: {D_69r(8^`x=~}. E DpED? SDC + Turbo Codes :9<;0= 9<>@?BADCE2GFIHKJ0LNM OQPR!STR!7VU3TW8RXPR!Y[Z78\\8W^]RX_ ` `badc30]\nef\hgji `lk$7)mfyl8\ief\inn; op7)mfq<\y[3t\\+3tr!qrrsu/8 7tBmf6<\\3umfqvPwxmfqq<8 S4mf]@`;y` `=Bc&8 7-;{z@\+30q5[]+w<8BZ7R!ZR \8W} ~ A@;.

36 UMTS channel simulations ber QCIF silent color video compressed to 200 kbits/s (30 frames/s). Y U V BER BER BER Channel Method Indoor UMTS SDC+TC Pedestrian UMTS SDC+TC Page 36 of 53 Vehicular UMTS SDC+TC Mean PSNR (db) results.

37 UMTS channel simulations ber! "$#%'&()+*,$# *-.&(,'0/1&(- 243!"$#% &$')( *+#% $#$,-.*,/#01 QCIF silent color video compressed to 200 kbits/s (30 frames/s) Page 37 of :(;<'=>@?BA.CEDGFIHJ5LKNM O9PRQTSVU-W!XYW!;[Z.6Y\(<-WU-W!]_^(;<-``<-\badce?4f4f0gih(69c`Rjd`lkm34fon;)ad]o<-ǸjdǸpR> qr;)ads`]_6y``6yw!stw ZB<+;[uEvwqxy7<-\(<-`c;6zadswU){ ads(s<+x7adc[f >@f4f }~h%<+;>ue`69s(8 x. ~Fƒ EqFE> SDC + Turbo Codes ;:=<7>@?BADC 2FEHGJI0KML NPOQ!RSQ!8UT3SV9QWOQ!XZY89[[9V]\ ^ <7_7_a`cb 30^[Md [fehg _ji 8)\ Xj9[Hd [Hk : lm8)\ n;[xz3s[[+3sq!noqpt/9 8Uq@roltsWu49V9[+^+8+3v\ n]owx\ nn;9 R\ ^ _:=_7_yBb 9 8 : q@[+30n5z^+w;9by8q!yq7[9v r{z A@:

$6WV2UX7S&QS YRZ67[([7X]\_^`> a$ $a&bcz;[de< agf#6(\_yo7[ih_[ijm:$ $7 6g\_mnDomp^7 6*m;7$ $6*2/m4uf#6S Ywv]xyZz\ Q{.$ ;> a:=v arx ˆ} ƒo ~?

$\T](;<-^)^<-Z_a`cb4d4d[egf(]^ghi?$ $op;)_aq^\t6w^^6wu!$

38 Internet channel simulations - 5 % packet loss QCIF Foreman color video compressed at 200 Kbps (30 frames/s)! #"$ %#&(' )*"$ #"#+,-)+."/0! "$#%'&()+*,$# *-.&(,'0/1&(- 243 Page 38 of 53 SDC. 132/ ;:=< >?A@CBED1GFIH JKMLON PRQS TUS 6WV2UX7S&QS YRZ67[([7X]\_^`> a a&bcz;[de< agf#6(\_yo7[ih_[ijm: kl6(\_m;[yr2u[[*2us m-s,v.7 6g\_mnDomp^7 6*m;7 ^g[*2/yq5tr\_^s 6C[*5st7 6*2/m4uf#6S Ywv]xyZz\ Q{.7 ^TUS [[ :~}[*2/m4 B: ƒo ~?;> a:=v arx ˆ} ƒo ~?z< :=Š_8uX ˆ 9ƒo ~?8a:=Œ >RX : :(;<=>@?4? ACBEDGF5IHKJ!LMONQP!RTS-U!VWU!;YX.6WZ(<-U[S-U!\T](;<-^)^<-Z_a`cb4d4d[egf(]^ghi?4dkj;)_a\Q<-^mla^mnO> op;)_aq^\t6w^^6wu!quxr<+;k_aqsftqu`<+;q<+`k^69\':(vv_a`u!;e^: w<+;69q(80j;u!\yxz{] _!S}r<+`gVWU4^^ >~ ^69q(8 ` < ](;U!]U4^<-Z ƒ kd >7 tŷ.š 'AŒb4 >$Ž! sz( g Œ~ tŷ.š 'A?42 >@24bsZ( g tŷ.š 'A7= d >@x4b Z( g>

39 Conclusions Redundancy parameter estimation uses channel characteristics; The MDBA automatically adapt the amount redundancy to channel characteristics; MDBA extension to video by adding to the 2D DWT a 1D DWT in the time direction. The 3D scan-based DWT transform allows automatic adaptation of the coding process to time varying states. The proposed MDBA is suitable for video transmission over time-varying channels. Page 39 of 53

40 5. Outline Quincunx images; MDBA for quincunx images; Taking account of CCD captors redundancies; Proposed criterion; Satellite simulator; Results; Conclusions. Page 40 of 53

41 Quincunx images Combining a pair of CCD linear arrays in a quincunx arrangement to improve image resolution. Each CCD linear array generates an image sampled on a square grid. CCD 1 CCD 2 CCD direction: 0.5 pixel Page 41 of 53 Velocity direction: n+0.5 pixel Figure 1: Representation of the two CCD linear arrays of a SPOT5 type acquisition system. Quincunx image Figure 2: Combination of a pair of CCD linear arrays in a quincunx arrangement

42 ! "$#%& ')(*%+',-.(0/213')40.(5%6(078.9: 6/"$#6 ;< #&9AB##')( C$DFEAG"HBI+JLK.EDFM-IONPK.Q&RSQ&R ITQ&U$DFV CXWZYU&E[G"UI+CIVQI+JLK.V]\_^ `0ab^ a cfd3egf R IOUI+C&h"H.Q&K.V"M N<D)i-IHBIQkj+Y,I=ljKBIVmQCnDFUIoh CI+JpK.VTQ&R I*qsrutwv]G"UI+CIVmQI+JTK.VOj$RxDFG"QIUny eoz K.M h"ui*^ e { G"UI+CIVQC2Q&R I Cj$R IE*I YFWQ&R I C&Q$DFV J DFUJ}qsr ~WYU2 h"k.v jh"v0 K.EDFM-I+C e MDBA for quincunx images Merge MD Bit Allocation 2 Coder Channel 1 Coder Channel 2 CHANNEL Decoder Page 42 of 53 Description 1 Central description Description 2 (R 1, D 1) (R 0, D 0) (R 2, D 2) z K.M h"uiƒ^ e { qsr ~WYU2 h"k.v jh"v0 K.EDFM-I+C e? Q$DFV J DFUJ q r ~ E*IQ&R Y,J e

43 D5EGFIHJF'KMLONQPSRTNQNJRTEMLUEMDV=WORTLOXYWOLZ[RPSK]\M^0NY<I_`F'Eba c EMNQ^0dePS^fHQgOEhd ikjmln.oprq sut.v2w#x!vzyptvblnvb{tv5t#y.jt }j~2os} lnvblx!sƒ G b }{{ 5#prlTs {Tv $~Tn {Ts$.ypTljv5pTpˆ s${tv!š.{tv5~2opv2w%: Œlnvzt#oŽ 'v2{tv2j l1t# $t#o~!o%q } :v5p! }{Tv!ypv5tƒlTs :v2jv2{ }ltvzlnvblxbst#ož 'v2{tv2j l t.v5pt~2{o%š.los$jp o%j p~tnv2q vm }jt lnvƒt#ož v2{tv2j~5vms} 0lnv5pTvƒlxbs o%q } :v5p op0 s$o%jv5t lts 's$ln t.v5pt~2{o%š.los$jpˆ1š v2l]ypm~ }w%w }jt lnv=lxbs t#ož v2{tv2j l= G b I }{{ up }jt P 1 P 2 Taking account lnv2o%{gt#ož 'v2{tv2j~5vœˆ1œ pz~ }j6 v ptv5v2j6o%j Œy.{Tvž#ˆŸž# #s$jv=t.v5pt~2{o%š.los$j ~5s$j l }o%jp ɛ of CCD captors redundancies }jt P 1 ɛ #xgn.o%wvhlnvƒs$lnv2{]~5s$j l }o%jp P 2 }jt ɛˆ CCD1 CCD2 MD Bit Allocation 3 3 P1 + ɛ P2 Coder Channel 1 Coder Residual Coder Channel 2 Multiplexing Multiplexing CHANNEL De multiplexing De multiplexing ˆP 1 ˆɛ 1 ˆP2 ˆɛ 2 Page 43 of 53 Decoder Description 1 Central description Description 2 (R 1, D 1) (R 0, D 0) (R 2, D 2) fo% Œy.{Tvƒž#ˆŸž# 1 h s${m y.o%j~2y.j# 6o%q } :v5p(ˆ1 0{Ts$Š'sŒpTv5t h q v2lnsut ˆ

44 Proposed Criterion J ({q i,1, q i,2, q i,ɛ }) = D λ j (R j R T /2) j=1 Page 44 of 53 We propose D 0 = N w i σi,0d 2 i,0 ( q i,1, q i,2, q i,ɛ ), i=1 D i,0 ( q i,1, q i,2, q i,ɛ ) = 1 [( σ 2 σi,0 2 i,1 D i,1 ( q i,1 ) + σi,2d 2 i,2 ( q i,2 ) ) ] + σi,ɛ 2 2r N D i,ɛ ( q i,ɛ ) 1 + r N Total side distortion for channel j D j = N [ σ 2 i,j D i,j + (σi,jd 2 i,j + σi,ɛd 2 i,ɛ ) ], j = 1, 2. i=1

45 "*( /=)O# *( '^O( "0='OP P O0# P_?=d- PSO ";- ' O# =)*(P^!&(Al?. =# = (i =.`pr ( U OP P O='# P 'oy&n)1- O#O "a]o9 \O# \p2p P6 /j MDC for A Quincunx <ca=-odo# ( 'OP P OcA=P=A-LOh# P %&'($ 'oy&n*) j?p>#. # O#Ovp2dP Z OM/&Pkp 23 ( 52 - O# l-oko'ookl5l%b( l!(al.d6o =0/Q*](cO'O; /&P=xO'O=g#= P=)*(P 6# P>TO#X/Q]*( O'O9-_ /&PcqO'Ocg#= =)](PxL-'OP>.,( l!co# :=)L-( O'OO`pRaO# )OAp>O# )P Z OV/&Pj g co O (&t4q Z # =)*(P< /Q 'P Otpfh0 P l%oo K=)*(PO O (K ^O k FpR Satellite simulator As satellite channel simulator we use the model proposed by O'O; ( N k Chee and Sweeney ;O'OH';) for LEO j O# =)*(P@-QO# 9 ^O Z P P/&O# ( O l-o l!o=5o# dc( l%(fl?.`o# b-= satellite p k ;2 % channels, ]OSP- 6# P)O# that b is threegood state, single error state Frichman N k l-o0( O l-o,0( l!(\l.o# d-=mp ;2*! ^OSPCj9p># d( Z ( model. O' O,ll P O4p29O# ao#a/q]*(.o'oqg O'#= 9=)*(PL$l% p2- ( 'oy&n*) j <9# U + /&-Mj O# 9p>#X/Q]*( O'O9-^ /&P6"O'O6g O#= =)*(PgpR 40 'j Page 45 of gb /&-Mj s p>#x/q*]($o'o9- /&PaTO'Oqg O#= 9=)](P@pR 40 'j

46 Simulations Proposed method: We use two pixels dyadic Nimes image; For spatial decomposition, the coder uses 9-7 biorthogonal filter and performs three levels of decomposition. Page 46 of 53 MDBA method: We use a pixels quincunx Nimes image generated interleaving by two dyadic image; For spatial decomposition, the coder uses (6,2) nonseparable lifting scheme and performs seven levels of decomposition.

47 Simulations - MDBA vs. quincunx images Side PSNR proposed MDC for Central PSNR Page 47 of 53 2 bpp Standard method Proposed method bpp Standard method Proposed method Table 1: PSNR values for Nimes image when considering transmission at an elevation angle of 40 ( ber). Side PSNR Central PSNR 2 bpp Standard method Proposed method bpp Standard method Proposed method Table 2: PSNR values for Nimes image when considering transmission at an elevation angle of 30 (0.001 ber).

Side Description Quincunx image compressed to 2bpp. Transmission: 0.001 ber Ln!

48 Side Description Quincunx image compressed to 2bpp. Transmission: ber Ln!" %$ %? $ #!! ## / Ly Page 48 of 53 gb /&- s * :.+(?= /QTA=( Odn<l 6# b ( /co' gb /&- C=O s * :.1(>?= /QcA=(,OWnl 6# t ( /bo' C=' $P Z O,/&Pq 30 6 al!+7 =i]o' ( ( 7=)O# ]( =i](a(]( OT$P Z O,/&Pq 30 6 al!+7 x%&'( 7=)O# *( =i](q(*( MDBA New MDC

49 Central Description!" %$ %? $ gb /&- Ln s #!! ## * :.+(%?= /Q"A=(dO?n9l / 6# k ( /0O' C=O Quincunx image compressedbp toz O./&PT 2bpp. 30Transmission: ber 6?l%+7 i]o' ( ( =)O# ]( #"]O'P%'# P Page 49 of 53 gb /&- Ln s * :.+(%?= /Q"A=(dO?n9l 6# k ( /0O' gb /&- Ly s C=O bp Z O./&PT 30 6 * :.+(%?= /Q"A=(dO?n9l 6# k ( /0O' C=O?l%+7 i]o' ( ( =)O# ]( #"]O'P%'# P Z OW/&P? P l!+7 x%&'()2=)o# ]( #"]O'P%'# P MDBA New MDC

50 Conclusions We proposed a MDC designed to get the best image quality after transmission over satellite channel; The proposed MDC take into account the redundancy between the two CCD; We adapted the MDBA for quincunx images. The proposed MDC for quincunx images over performs the MDBA for high levels of noisy channel. Page 50 of 53

51 Page 51 of General conclusions and perspectives General conclusions We proposed a new MDC based on bit allocation - MDBA; The bit allocation algorithm dispatches the redundancy between the descriptions; MDBA can adapt automatically to time varying networks; MDBA show good performances for different transmission channels involving image or video: Gaussian, UMTS, Internet. We present an adaptation to quincunx images and applied it for satellite channels; MDBA is comparable to the SDC when noiseless channels; MDBA performances are comparable or better than other MDC; MDBA is comparable to Turbo Codes for stationary channels;

52 Perspectives Generalization of the MDBA to N descriptions; How to dispatch redundancy in this case? Consider motion compensation for video; Page 52 of 53 Take into account a real UMTS channel model; Video streaming.

Figure 3: Lena image coded at 1.0 bpp. BSC channel at 0.001 ber. PSNR=33.89 db.

53 Figure 3: Lena image coded at 1.0 bpp. BSC channel at ber. PSNR=33.89 db. Figure 4: Lena image coded at 1.0 bpp. Gaussian channel at ber. PSNR=34.90 db. Page 53 of 53 Figure 5: Original Lena image.

Multiple Description Coding for quincunx images.

Multiple Description Coding for quincunx images. Application to satellite transmission. Manuela Pereira, Annabelle Gouze, Marc Antonini and Michel Barlaud { pereira, gouze, am, barlaud }@i3s.unice.fr I3S