Download Multiple Description Coding for Error Resilient Multimedia Communication - Prof. Min Wu and more Study notes Electrical and Electronics Engineering in PDF only on Docsity!
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Multiple Description Coding Multiple Description Coding
Min Wu
Electrical & Computer Engineering
Univ. of Maryland, College Park
http://www.ece.umd.edu/class/enee739m/
[email protected]
ENEE739M Spring 2002^ ENEE739M Spring 2002LectureLecture-
-13^13
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [2]
Review of Last ClassReview of Last Class "^
Fine granularity scalable coding (contíd)^ ñ
Performance ñ^
Discussions
"
Wavelet coding for achieving scalability^ ñ
Multi-resolution nature of wavelet transform ñ^
FGS through wavelet tree structure and bit-plane coding
"
Today^ ñ
A related problem to scalable coding and error resilient coding^ #
Multiple Description Coding
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [3]
Error Resilience via Layered Coding Error Resilience via Layered Coding "^
(recall) Layered coding for achieving scalability^ ñ
Base layer (low but acceptable quality) and enhancement layers
"
Unequal error protection^ ñ
Protect base layer strongly to ensure it is delivered essentially error-free
#^
e.g., via strong ECC coding
ñ^
Enhancement layers only use base-layer frame as prediction ref.^ #
avoid error propagation at expense of coding efficiency
"
What if a channel may temporarily go down so that baselayer may not be deliverable through that channel?^ ñ
Packet-loss channels such as the Internetñ Wireless channels with severe fading from time to timeñ Possible to have access to multiple independent channels
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [4]
Class Discussion on Last QFT^ Class Discussion on Last QFT
"
Transmit an image (say, Lena) using two channels^ ñ
Data sent over each channel may get lost ñ^
Conditions of two channels are independent^ #
Prob. that both channels are bad at thesame time is very small
[ Codec Design Goal ] $^
Fix the total # of bits to be sent over the two channels $^
When only one channel is good, want to get imagedecoded with reasonable perceptual quality $^
When both ch. are good, want to get image decodedwith improved quality
Diagram is from Wangísbook Fig. 14.
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [6]
Multiple Description CodingMultiple Description Coding
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [7]
Multiple Descriptions Coding (MDC)Multiple Descriptions Coding (MDC) "^
What is MDC?^ ñ
Several bitstream (ìdescriptionsî) of the same source are sent overseparate channels that are often of equal quality ñ^
Reconstructable through subsets of descriptions ñ^
Higher quality when having more descriptions at receiver ñ^
Simplified scenarios: two balanced descriptions ñ^
Info. theoretical aspects back to early 1980s
"
Basic ideas of MDC^ ñ
Introduce redundancy/correlation to different descriptions to helpestimate lost descriptions^ #
simplest but inefficient method: send same description over each ch.
ñ^
In contrast, layered coding often uses hierarchical, decorrelatingdecomposition ñ^
MDC is more useful than SDC+ER/EC when ch. loss is likely to occur
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [8]
Typical Approaches to Realize MDCTypical Approaches to Realize MDC "^
Interleaved temporal sampling^ ñ
ìVideo redundancy codingî in H.
"
Multiple Description Scalar Quantizer^ ñ
(two-description case) One central quantizer and two side quantizers
"
Pairwise Correlating Transform^ ñ
Introduce controlled correlation to two originally uncorrelated r.v.
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [9]
Temporal Video Redundancy Coding (VRC)Temporal Video Redundancy Coding (VRC) "^
Basic idea of VRC^ ñ
Improve temporal error resilience through multiple prediction ñ^
Divide picture sequence into several threads and assign pictures to one ofthe threads in a round-robin fashion ñ^
If some threads are damaged, surviving threads can be used to predict thenext sync frame and conceal frames in damaged threads
"
Resilience performance^ ñ
Claimed good video quality at frame loss rate up to 20% via 3-thread 3-picture-per-thread VRC
"
Cons^ ñ
Substantially reduced coding efficiency due to penalty in ME/MC formore separated frames in one thread
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [14]
MultipleMultiple-
-Description ScalarDescription Scalar Quantizer
Quantizer (contíd)
(contíd)
"
Optimization framework for optimum codec design^ ñ
For specific 1-ch. coding rate and 2-ch. coding rate ñ^
To find optimum MDSQ codec to minimize 2-channel reconstruction errorsubject to 1-channel reconstruction errors below thresholds ñ^
Lagrangian multiplier approach ( Sec.III of Vinayí93 )
"
Iterative design algorithm^ ñ
(simplify the design) Consider a specific choice of index pairing andassignment^ #
prefer closer to main diagonal to avoid excessive single ch.reconstruction error ~ minimize ìspreadî of each recon. Value
ñ^
(decoder refinement step) Given quantizersí partitions, determine optimumreconstruction points for both 2-ch. and 1-ch. Cases ñ^
(encoder refinement step) Refine quantizersís partitions ñ^
Simplification in iterative designs for squared-error distortion measures
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [15]
MDC UsingMDC Using Pairwise
Pairwise Correlating Transform (PCT)
Correlating Transform (PCT)
"
Basic idea^ ñ
Generate two descriptions that are uncorrelated within same description butpairwise correlated across different descriptions ñ^
Introduce correlation to two quantized uncorrelated r.v. ñ^
Encode two correlated r.v. from PCT output ñ^
Perform inverse PCT when both r.v. are received ñ^
Estimate the missing r.v. then perform inverse PCT if only one is recíd
"
Redundancy-Rate Distortion (RRD)^ ñ
Redundancy rate ~ rho( D1; D0 )^ #
Given two-ch. distortion D0 (or the 2-ch. rate), how many bits ofredundancy per r.v. to achieve a desired 1-ch. distortion D1 are needed?
ñ^
rho ranges from 0 in single description case to R(D0) when sending samedescriptions over both channel^ #
D1 from (sigma^2 + D0)/2 to D0 where sigma is source r.v.ís variance
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [16]
Issues in Designing PCTIssues in Designing PCT-
-based MDCbased MDC
"
Optimal PCT
"
Integer-to-integer PCT on quantized coefficient pairs
"
Optimal redundancy allocation among different pairs
"
Optimal pairing strategy under fine vs. coarse quantization
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [18]
SummarySummary "^
Multiple Description Coding^ ñ
Conceptual comparison with layered coding ñ^
MDC via temporal interleaving ñ^
MDC via multiple description scalar quantizer ñ^
MDC via pairwise correlating transform
"
Next Time:^ ñ
Performance of MDC ñ^
Issues on error resilient communications
M. Wu: ENEE739M Multimedia Comm. & Info. Security (S'02)
Lec13 ñ MDC 3/19/02 [19]
Suggested Reading Suggested Reading "^
Multiple description coding^ ñ
Sec.8.4.8 of Sun-Reibmanís bookñ Sec.14.4.5 of Wangís video textbookñ Yao Wang et al. MDC paper via pairwise correlating transf. in IEEETrans. on Image Proc. 3/01 pp351-366ñ Vinay Vaishampayanís MDSQ paper in IEEE Trans. on Info. Theory5/93 pp821-
