Zero-Delay Multiple-Descriptions Source Coding of Stationary AR(p) Sources Using Feedback
Author
Søgaard, Kristian
Term
4. semester
Education
Publication year
2022
Submitted on
2022-06-03
Pages
102
Abstract
This thesis studies zero-delay multiple-descriptions (ZDMD) coding for stationary scalar AR(p) sources under mean squared error (MSE) constraints. AR(p) refers to signals where each sample can be predicted from the previous p samples. Multiple-descriptions coding splits data into several descriptions; in the symmetric two-description case, either packet alone supports a coarse reconstruction, while receiving both yields higher fidelity. Zero-delay means encoding and decoding sample-by-sample without buffering delay. The thesis introduces a new operational ZDMD scheme that combines index assignment with differential pulse code modulation (DPCM) quantization and assumes feedback from the decoder. The method quantizes the prediction error and maps it into two descriptions using a non-linear one-to-many mapping. The decoder reconstructs from whatever packets arrive; quality improves when both packets are received. The scheme is applicable regardless of the source model order p. In the DPCM stage, minimum MSE (MMSE) predictors—predictors that minimize average squared error—are used. Theoretical performance is analyzed, and practical performance is demonstrated through simulations. Results show that the proposed ZDMD scheme achieves lower MSE than existing ZDMD coding methods and is robust to packet losses.
Denne afhandling undersøger nul-forsinkelse multiple-descriptions (ZDMD) kodning af stationære skalære AR(p)-kilder under middelkvadratfejl (MSE) begrænsninger. AR(p) betyder, at hver værdi kan forudsiges ud fra de p foregående værdier. Multiple-descriptions kodning deler data i flere beskrivelser; i den symmetriske to-beskrivelser-variant kan hver pakke alene give en grov rekonstruktion, mens begge sammen giver højere kvalitet. Nul-forsinkelse betyder, at kodning og dekodning sker uden bufferingsforsinkelse, én prøve ad gangen. Afhandlingen præsenterer en ny operationel ZDMD-metode, der kombinerer indeks-tilordning og DPCM-kvantisering (differential pulse code modulation) og antager feedback fra dekoderen. Metoden kvantiserer forudsigelsesfejlen og afbilder den til to beskrivelser via en ikke-lineær én-til-mange afbildning. Dekoderen rekonstruerer ud fra de pakker, der modtages; nøjagtigheden er højere, når begge pakker når frem. Skemaet kan anvendes uanset kildens modelorden p. I DPCM-delen bruges MMSE-prediktorer, dvs. forudsigere der minimerer den gennemsnitlige kvadratiske fejl. Den teoretiske ydeevne vurderes, og den praktiske ydeevne demonstreres i simulationsstudier. Resultaterne viser, at den nye ZDMD-metode opnår lavere MSE end eksisterende ZDMD-kodningsmetoder og udviser robusthed over for pakketab.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords