Digital Signal Processing Reference
In-Depth Information
To estimate the error-propagation effect in future frames, the error-propagation count
“EPC(
n,m
)” of the Packet
P
n,m
is used to approximate the motion reference ratio in the
current GOP:
(
θ
-
n
)
(
θ
-
n
)
n
n
Knm
( ,)
Knm
(2,)
last
last
(3)
.
n
n
EPC( ,
nm
)
=⋅
β
α
⋅
inter
+⋅
β
α
⋅
inter
1
m
2
m
K
K
θ
=+
n
1
θ
=+
n
1
where “
n
last
” denotes the last frame in the current GOP that contains the Packet
P
n,m
;
“
” and “
” (
) are two weighting parameters whose values can be
β
β
β+=
1
1
2
1
2
empirically set to “
”. The error-propagation effect among
future frames is approximated as a frame-by-frame decaying process. After encoding
the Frame
n
, “
GTD
(
n
,
m
)” denotes the normalized GOP-level transmission distortion
induced by the lost Packet
P
n,m
,
where only Packet
P
n,m
is lost and other packets in the
same GOP are error-free.
” and “
β =
0.8
β
=
0.2
1
2
]
[
Dnm
(, ) EPC(, ) 1
⋅
nm
+
(4)
i
.
GTD n m
(, )
=
S
nm
,
Based on the Equation (4), the propagation-aware online packet classification
assumes that the Packet
P
n,m
with the larger
GTD
(
n
,
m
) value is more important, and is
classified as the high-reliability traffic class. According to one of the aforementioned
three online packet classification schemes, the sender can timely determine the traffic
class for each packet in the current frame.
4
Simulation Results
By using H.264 reference software JM17.0 [7], three typical CIF (4:2:0) sequences,
namely
Akiyo
,
Foreman
and
Football
, are encoded at their typical bitrates. The GOP
structure is an I-frame followed by a serial of P-frames. The default error concealment
in JM17.0 decoder is performed. After each frame is encoded, the most important
packets are classified as the high-reliability traffic class, while other packets in the
same frame are classified as the best-effort traffic class.
Without loss of generality,
the high-reliability packet loss probability
p
h
is equal to 0%, and the best-effort packet
loss probability
p
b
varies from 0% to 20%. For each frame, the amount of high-
reliability packets “
η
” includes four typical values from low (
η
=1) to high (
η
=4) so
as to highlight the characteristics of the proposed scheme. For its simplicity and
mathematical tractability, the two-state Markov channel model has been widely used
to generate different packet loss patterns for the Internet. Before each simulation run,
the channel model is reinitialized by a time-based seed.
We run 10 simulations with
each
η
value, and calculate the peak signal-to-noise ratio (PSNR) of each received
frame. For a given
p
b
value and a test sequence,
“PSNR(
p
b
)” denotes the average
PSNR over 40 different simulations with four typical
η
values. The main simulation
parameters are listed in Table 1.
