TITLE: Joint Denoising and Decoding of Known/Unknown
Sources
SPEAKER: Krishnamurthy Viswanathan, HPL and UCSD
DATE: 3:00 - 4:00 P.M., Friday September 5, 2003
LOCATION: Half Dome, 3L (PA)
HOST: Vinay Deolalikar
ABSTRACT: We consider the setting where an
uncompressed source stream is encoded via a channel code for transmission over a
noisy channel. In this setting we seek to improve the decoding performance by
taking advantage of the inherent source redundancy, both when the source is
known and in a universal setting. In these investigations we employ
Repeat-Accumulate (RA) codes, which offer very good performance at
low-complexity.
Three types of joint denoising and decoding were considered: hard denoising
followed by hard-input decoding, soft denoising followed by soft-input decoding
and an iterative denoising-decoding procedure. For the case of a known Markov
source, the structure of the joint denoiser-decoder is expressed in terms of a
factor graph over which a belief propagation algorithm is performed. This
amounts to combining the forward-backward algorithm with conventional belief
propagation for channel codes. Significant improvements were observed over
simple decoding. These results were nearly replicated when the source is assumed
to be unknown and the forward-backward denoising algorithm is replaced with a
Discrete Universal Denoiser(DUDE). The performance of the universal joint
denoising-decoding algorithms on images was also studied.
The talk will include an overview of concepts related to graph-based codes
such as factor graphs and belief propagation decoding.
|
printable version
