DATE: Tuesday, March 31
TIME: 4 p.m.
LOCATION: Beckman Room 5602

SPEAKER: 
                              Prof. Edward J. Delp
                   Video and Image Processing Laboratory
                         Purdue Multimedia Testbed
               School of Electrical and Computer Engineering
                             Purdue University
                          West Lafayette, Indiana

TITLE: Error Concealment in Encoded Video Streams

In ATM or wireless networks, cell or packet loss due to channel errors or
congestion causes data to be dropped in the channel which results in the
loss of entire macroblocks when MPEG compressed video is being transmitted.
Two issues need to be addressed to alleviate this problem: finding the
location of the missing data and processing the sequence to recover the
missing data. Two approaches have been used: active concealment and passive
concealment.  In active concealment error control coding techniques are
used along with retransmission. In passive concealment the video stream is
post-processed to reconstruct the missing data.  Passive concealment is
necessary in many applications where error control coding cannot be used
due to compliance with video transmission standards or when active
concealment fails.

We will describe several techniques we have developed for the detection and
reconstruction of missing data using passive concealment. We have addressed
this problem to date in the context of concealing errors  in MPEG encoded
streams transmitted over ATM networks.

We will describe techniques for packing ATM cells with compressed video
data, with the aim of detecting the location of missing MPEG macroblocks in
the encoded video stream. This technique also permits proper decoding of
correctly received macroblocks,  and thus prevents the loss of data from
affecting the decoding process. We have also developed passive spatial and
temporal techniques for the recovery of lost macroblocks in a compressed
video stream.  The spatial techniques fall into  two categories:
deterministic and statistical.  A deterministic spatial approach aims at
reconstructing each lost pixel by spatial interpolation from the nearest
undamaged pixels.  In the statistical approach, each frame is modeled as a
Markov Random Field, and a maximum a posteriori (MAP) estimate is obtained.

In temporal reconstruction, a search is carried out over a reference frame
for the macroblock sized region that will maximize the posterior
distribution of the lost macroblock given its neighbors.  We have also
developed a fast sub-optimal temporal estimate that searches for motion
vectors that yield the MAP estimate for only the boundary pixels.