TITLE:      POCS based error concealment for packet video
AUTHORS:    G. Yu and M. W. Marcellin
CONFERENCE: 1997 Data Compression Conference, Snowbird, Utah, March 1997
ABSTRACT:
This paper proposes a new error concealment algorithm for packet video,
effectively eliminating error propagation effects. Most standard video CODECs
use motion compensation to remove temporal redundancy.  With such motion
compensated interframe processing, any packet loss may generate serious error
propagation over more than 10 consecutive frames. This kind of error
propagation leads to perceptually annoying artifacts. Thus, proper error
concealment algorithms need to be used to reduce this effect.
 
The proposed algorithm adopts a one pixel block overlap coding structure to
solve the error propagation problem. If no packet loss occurs, the decoded
pixel intensities on the overlap areas should be consistent (with small
differences caused by quantization error.) When a packet loss occurs, the
corresponding reconstructed frame and any frames referring to it are all
damaged. Such damage causes inconsistent pixel intensities on the overlap
areas of damaged frames.
  
The proposed error concealment method poses the packet loss recovery problem
as one of parameter estimation.  Lost transform coefficients are estimated by
the method of Projection Onto Convex Sets (POCS).  The estimation is performed
in a manner that maximizes the consistency of pixel intensities in the overlap
areas of the reconstructed frames.
   
There are three types of convex sets used in the proposed algorithm. One is
the constraint on the "non-lost" transform coefficients. Another is the
constraint on the consistency of pixel intensities on the overlap areas. The
last convex set is the constraint on some pixels which have equal intensities.
The proposed algorithm needs to alternatively project onto these sets until
reaching convergence.
    
The proposed POCS based method has consistently high performance in
error concealment. Experimental results (using a modified version of CCITT
H.261) show that it can have good error concealment results even when the
damaged frame loses all DCT coefficients.