Page 246 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 9.7. Postprocessing (or Concealment) Techniques 223

temporal replacement (TR) because it e ectively replaces the damaged block
by its corresponding block in the reference frame. This method works well
for stationary and quasi-stationary areas, e.g., background, but will fail for
fast-moving areas.
Another technique is to exploit the high-correlation property of video block-
motion elds and replace the damaged motion vector with the average (AV)
[179, 190, 189, 191, 192] or the median [179, 192] of neighboring vectors.
This technique works well for areas with smooth motion but will fail for areas
with unsmooth motion, e.g., at the boundaries of objects moving in di erent
directions.
A boundary matching (BM) technique has also been used to select a suitable
replacement from a set of candidate motion vectors [196, 197, 198]. Assume
that a set of M neighboring motion vectors V={v 1 ; v 2 ;:::; v M } is to be used
for the concealment of a damaged block D of size N × N with its top-left
y
x
corner at (x o ;y o ). Each candidate vector v i =(v ;v )in V is used to conceal
i
i
the damaged block D. The quality of this concealment is assessed using the
continuity across the concealed block boundaries. This continuity is measured
using the side-match distortion (SMD) measure, de ned as
R
L
B
T
SMD i = SMD + SMD + SMD + SMD ; (9.1)
i
i
i
i
L
where SMD is the sum of absolute, or squared, di erences across the left
i
boundary of block D when concealed using candidate vector v i . Thus
�
N−1
y
L
x
SMD = g[f t (x o − 1;y o + k) − f t−Ot (x o + v ;y o + v + k)]; (9.2)
i
i
i
k=0
where f t and f t−Ot are the current and reference frames, respectively, g =(·) 2
T
R
for the SSD, and g = |·| for the SAD. Similarly, SMD ; SMD and SMD i B
i
i
are the side-match distortions across the right, top, and bottom boundaries,
respectively. Based on the smoothness property of video signals, the candidate
motion vector that achieves the minimum SMD is chosen as the recovered
motion vector. Thus
v ˆ = arg min SMD i : (9.3)
v i ∈V
The main advantage of this method is that displacement estimation is based on
a distortion measure. The method will fail for areas with unsmooth motion and
also for areas with low spatial correlation, e.g., at the boundaries of objects.
Similar to spatial concealment, Bayesian statistical approaches have also
been used for motion vector recovery, e.g., Ref. 195.