A concealment method for video communications in an error-prone environment

Abstract

In this paper, we propose a two-stage error-concealment method for
block-based compressed video which was transmitted in an error-prone
environment. In the first stage, we obtain initial estimates of the
missing blocks. If the motion vectors associated with the missing blocks
are available, motion compensation is used to provide good estimates.
Otherwise, a novel algorithm which preserves image continuity is used to
estimate the blocks. In the second stage, a maximum a posteriori (MAP)
estimator, which employs an adaptive Markov random field (MRF) as the
image a priori model is used to improve the video reconstruction
quality. The adaptive model enables the estimation to incorporate
information embedded not only in the immediate neighborhood pixels but
also in a wider neighborhood into the reconstruction procedure without
increasing the order of the MRF model. The proposed concealment method
achieves very good computation-performance tradeoffs, as demonstrated
via experimental results

Full text

1122 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 18, NO. 6, JUNE 2000
A Concealment Method for Video Communications
in an Error-Prone Environment
Shahram Shirani, Member, IEEE, Faouzi Kossentini, Senior Member, IEEE, and Rabab Ward, Fellow, IEEE
Abstract—In this paper, we propose a two-stage error-conceal-
ment method for block-based compressed video which was trans-
mitted in an error-prone environment. In the first stage, we obtain
initial estimates of the missing blocks. If the motion vectors associ-
ated with the missing blocks are available, motion compensation
is used to provide good estimates. Otherwise, a novel algorithm
which preserves image continuity is used to estimate the blocks. In
the second stage, a maximum a posteriori (MAP) estimator, which
employs an adaptive Markov random field (MRF) as the image a
priori model, is used to improve the video reconstruction quality.
The adaptive model enables the estimation to incorporate infor-
mation embedded not only in the immediate neighborhood pixels
but also in a wider neighborhood into the reconstruction proce-
dure without increasing the order of the MRF model. The pro-
posed concealment method achieves very good computation–per-
formance tradeoffs, as demonstrated via experimental results.
Index Terms—Error concealment, image reconstruction,
Markov random fields.
I. INTRODUCTION
DIGITAL image and video signals require very high bit
rates, thus, compression of such signals before their trans-
mission is necessary. Communication channels are not error
free and, consequently, the encoded bit streams are vulnerable
to transmission errors—usually causing loss of blocks of data
and/or loss of synchronization. Despite the various methods that
have been proposed to combat or localize the effects of channel
errors, the quality of a decoded video sequence may degrade
significantly because of the residual errors. Error concealment
intends to conceal the effects of such errors by exploiting redun-
dancies in the video signal and limitations of the human visual
system, without requiring additional information.
Temporal concealment methods are usually used for error
concealment in inter-coded frames. Data of an inter-coded block
are composed of the motion vector and the DCT coefficients of
the prediction error. If the motion vectors are received without
errors, the missing blocks are set to their corresponding mo-
tion compensated blocks. Since in many of the video compres-
sion algorithms (e.g., H.263) the coded motion vectors and the
DCT coefficients are interleaved, the loss of data of a block
usually results in the loss of both the motion vectors and the
Manuscript received May 4, 1999; revised November 30, 1999. This work
was supported by MDSI, BC ASI, and NSERC.
The authors are with the Department of Electrical and Computer Engineering,
University of British Columbia, Vancouver, BC, V6T 1Z4, Canada.
Publisher Item Identifier S 0733-8716(00)04337-7.
DCT coefficients.1Therefore, most of the proposed temporal
concealment methods first estimate the motion vector associ-
ated with a missing block using the motion vectors of adjacent
blocks [1], [2]. The estimated motion vector is then used to find
a block in the previous frame which yields the restoration of
the missing block. One problem associated with these methods
is that if the adjacent blocks are coded in a nonpredictive way
(e.g., intra-coded), there would not be any data available to esti-
mate the missing motion vectors. Thus, ad hoc assumptions for
estimating the missing motion vector should be made, and the
results may be unreliable. Moreover, if the missing block lies
on the boundary of two objects moving in opposite directions,
such methods will perform quite poorly.
Spatial error-concealment methods restore the missing
blocks using the information in the current frame. In [3],
to restore the missing data, a measure of variations (e.g.,
gradient or Laplacian) between adjacent pixels is minimized.
The underlying smoothness assumption of this method limits
its ability in restoring image details. In [4], each pixel in a
damaged block is interpolated from the corresponding pixels in
its four neighboring blocks such that the total squared border
error is minimized. In [5] and [6], the missing information
is interpolated utilizing spatially correlated edge information
from a large local neighborhood. Note that although these
edge-based methods are generally more accurate than other
approaches, they are computationally more intensive. In [7],
a computationally simple, spatial directional interpolation
scheme has been proposed. The two nearest surrounding
layers of pixels of a missing block are converted into a binary
pattern to reveal the local geometrical structure. Then, the
missing pixels are interpolated in a way to preserve the local
geometrical structures. In the statistical error concealment
methods, for example the one proposed in [8], it is assumed
that the pixel values in an image or video signal are realizations
of an underlying statistical model (e.g., MRF model). The sta-
tistical approaches of spatial error concealment are expected to
outperform the deterministic ones, as they provide a systematic
way for incorporating the a priori information about the video
signal in the restoration procedure.
Previous spatial error-concealment methods employing the
MRF model usually yield blurry images with a significant loss
of detail in the high frequency or edge portions of the image.
This is due to 1) the type of the MRF selected as the image
model, and 2) the fact that the amount of information that is
1Recognizing the need to improve the concealment capabilities, in the new
generation of video compression standards (e.g., MPEG 4, H.263++), the mo-
tion information is separated from other information through partitioning of the
coded data.
0733–8716/00$10.00 © 2000 IEEE

SHIRANI et al.: CONCEALMENT METHOD FOR VIDEO COMMUNICATIONS 1123
used in the reconstruction process is often restricted to a single-
pixel wide region around the erroneous area. Incorporating more
pixels in order to enhance restoration of edges and high-fre-
quency parts would usually require a higher order MRF model.
However, this is computationally expensive, as the complexity
grows exponentially with the order of the MRF model.
In this paper, we propose a two-stage error-concealment
method for compressed video. In the first stage, if applicable,
we use the information in the previous frame to obtain initial
estimates of the missing blocks. If the motion vectors of the
missing blocks are available, motion compensation is used to
provide the estimates. Otherwise, an algorithm which preserves
image continuity is employed to compute the initial estimates.
In the second stage, a MAP estimator is used for refinement of
the initial estimates. The MAP estimator employs an adaptive
MRF as the image a priori model. The proposed adaptive MRF
takes into account the local image characteristics embedded
not only in the immediate neighborhood pixels of the damaged
area but also in a wider neighborhood without a dramatic in-
crease in computational complexity. Our concealment method
improves on the existing statistical methods in that it yields
good reconstruction performance regardless of the content of
the missing blocks.
To make the compressed data more error resilient, most of
the standard-compliant video compression systems partitioned
a video frame into Groups of Blocks (GOB’s) or “slices,” which
are coded independently. Therefore, the output bit stream usu-
ally consists of segments separated with markers, where each
segment corresponds to the coded data of the blocks in a GOB or
a slice. When channel errors occur, the decoder usually discards
the erroneous data between two markers surrounding the erro-
neous data, effectively discarding the GOB or the slice. Then,
loss of data of a slice does not affect the rest of the compressed
video sequence. In our work, it is assumed that the frames of
the coded video sequence are partitioned into GOB’s or slices
and, thus, the missing data belong to blocks of a GOB or a slice.
Moreover, it is assumed that the decoder knows the locations of
the missing blocks. This information (e.g., the checksum infor-
mation) can be obtained from the network or it can be inferred,
for example, by detecting the semantic or syntactic violations as
a result of errors [9].
The rest of this paper is organized as follows. In Section II,
MAP estimation of missing data is briefly reviewed. Section III
presents the proposed error concealment method. Section IV
discusses the computational complexity of the proposed
method. Sections V and VI present our experimental results
and conclusions, respectively.
II. MAP ESTIMATION OF MISSING DATA
Let the corrupted image be represented by the vector , while
the decompressed image without errors can be represented by
. Using MAP estimation, the decompressed image estimate is
given by
Using the Bayes’ rule we, obtain
(1)
where the term has been dropped because it is inde-
pendent of . The corrupted image can be expressed as
, where is the transform that converts the decompressed
image into the corrupted one .2The conditional probability
of the corrupted image can be then written as follows:
(2)
Substituting (2) in (1), the MAP estimation becomes
(3)
where .
For image a priori model , we use the MRF.
According to the Hammersley–Clifford theorem, every MRF
should have a Gibbs distribution [10] in the following form:
(4)
where is a normalizing constant, is called a potential
function and is a function of a local group of pixels called
clique, and denotes the set of all cliques throughout the image.
The potential function characterizes the relationship between a
group of pixels by assigning larger costs to configurations of
pixels which are less likely to occur. The choice of the poten-
tial function impacts substantially the performance of the image
model. The function should be convex in order
to have an easily obtainable global minimum. Otherwise, local
minima will be present, and the function must then be mini-
mized via a computationally expensive technique such as simu-
lated annealing. Commonly, the potential functions are selected
to be of the form
(5)
where is a coefficient vector, is the vector of pixels in the
clique , and is the cost function. The convexity propertyof
the cost function insures that the remains convex.
The coefficients are usually selected such that provides
an approximation of the first or second derivative of the image
at each pixel. For the special case of , the model is
called a Gauss–Markov random field (GMRF).
III. PROPOSED METHOD
As mentioned earlier, our method consists of two stages. The
purpose of the first stage of our method is to obtain an initial es-
timate of the missing block using information from the previous
frame. If the missing block has been intra-coded, the initial esti-
mate is set to zero. If the missing block has been inter-coded, and
its motion vector is received correctly (e.g., because the encoded
2See [8] for a more detailed discussion of the transform.

1124 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 18, NO. 6, JUNE 2000
(a)
(b)
Fig. 1. (a) The four subblocks, (b) their corresponding subblocks in the
previous frame, and the blocks connected to them.
data have been partitioned), then the estimate is set to the corre-
sponding motion compensation block. When the motion vector
is not available, we find an estimate of the missing block such
that image continuity inside the block and across its boundaries
is preserved [11]. To do this, subblocks adjacent to the missing
block, i.e., , and in Fig. 1(a), are considered. First,
for each of these subblocks, a corresponding subblock in the pre-
vious frame is determined. The corresponding subblock is found
by searching a small area around the point corresponding to the
center of each of the subblocks in the previous
frame. The sum of absolute differences is used as the measure
of similarity. The four subblocks and their corre-
sponding subblocks in the previous frame are
shown in Fig. 1(a) and (b). For example, is the subblock cor-
responding to . Then, four blocks, namely , and
, which are connected to , and (respectively),
are determined. To obtain an initial estimate of the missing block
that smoothly connects to the rest of the image, a block from the
above four blocks that minimizes the squared sums of border er-
rors, between the estimated block and its adjacent above and left
blocks, is selected. Thus,
where
(6)
Each of the border errors and is defined in terms of pixels
by
and
Fig. 2. A pixel, its clique
c
, and the eight directions. The complement of the
clique
c
is the dark area.
where the vector consists of the bottom line pixels of the
block to the top of the missing block, and consists of the
right column pixels of the block to the left of missing block.
The vectors and are those elements of the estimated block
that correspond to the pixels in its top row and left column,
respectively. To ensure the online applicability (e.g., conceal-
ment while decoding the blocks in raster scan), only subblocks
to the left and above the missing block are used. If the condi-
tion of online concealment is relaxed, the same method can be
extended to include the subblocks of the blocks to the right and
below of the missing block.
The estimation of motion of the subblocks has a high compu-
tational overhead which can possibly introduce unacceptable re-
quirements on decoder. The computational overhead can be re-
duced if the search for displacement of each of the subblocks is
restricted to a set of candidate motion vectors. This is a decoder
option and can be used to trade performance against computa-
tional complexity. The set consists of the motion vector of the
block corresponding to the missing block in the previous frame,
the motion vectors of available neighboring blocks, the median
of the motion vectors of available neighboring blocks, the av-
erage of the motion vectors of available neighboring blocks, and
the zero motion vector [2].
In the second stage of our proposed error-concealment
method, the information around each missing block is used
to refine the initial estimate. This stage is based on a MAP
estimator. We consider a GMRF model with an eight-pixel
clique around each pixel as the image a priori model (see
Fig. 2). The reason for selecting an eight-pixel clique in the
manner shown in Fig. 2 will become clear later. The potential
function of (5) is selected as
(7)
where is the clique, is the weight assigned to the
difference between the values of the pixel in position and
the pixel in its clique at position and .
Combining the MAP estimator of (3) with the image model
[(4) and (7)], the restoration of missing data eventuates in the
following minimization problem:
(8)

SHIRANI et al.: CONCEALMENT METHOD FOR VIDEO COMMUNICATIONS 1125
where is the set of all missing pixels in the frame. Since
is a convex function, the above minimization
problem yields a unique global solution. In fact, the estimated
value of a pixel is given by
(9)
where is the clique and is its complement shown in Fig. 2.
In our adaptive MRF model, the weight corresponding to the
difference between a pixel and one of the pixels in its clique
[in (9)] is selected adaptively, based on the likelihood
of an edge in the direction of the subject pair of pixels. The ra-
tionale behind this selection is to weigh more the difference be-
tween the pixels in that direction which will cause the values
of the pixels in that direction to get closer to each other. The
likelihood of edges in each of the eight directions is computed
using blocks around the missing block. In this way, the available
information in a larger area is exploited without increasing the
order of the MRF model (which increases dramatically the com-
putational complexity). To determine the likelihood of edges in
each of the eight directions, edges in the blocks surrounding the
missing block whose directions imply that they pass through the
missing block are determined. That is, we first obtain
and
for every pixel in the blocks to the left, right, top, and bottom of
the missing block. The magnitude and angular direction of the
edge at pixel are
and
where determines if the edge at pixel passes through
the missing block. Since there are eight pixels in the clique,
the value of is rounded to one of the eight directions equally
spaced in the range from 0 to 180 . There is a counter
for each of the eight directions. If the ex-
tension of an edge at pixel belonging to one of the neigh-
boring blocks passes through the missing area, the counter for
that particular direction is incremented by the amount of .
Since the employed edge detector is sensitive to the image noise,
the values of s, are compared to a threshold. If
any of them is less than the threshold, it is set to zero.
There are eight pixels in the clique of each pixel, and eight di-
rections for the detected edges. Each pixel in the clique of a pixel
corresponds to a direction. In our proposed method,
is selected based on the edge counter of the direction corre-
sponding to and , i.e.,
(10)
where is a constant and is the counter corresponding to di-
rection , and direction corresponds to the direction formed
by and . Finally, the second stage of the proposed
error-concealment method can be summarized as follows. 1)
Determine the edges in the neighboring blocks and assign them
to eight equally spaced directions. Compute the counter for each
direction. 2) Use (10) to find a set of weights for each missing
block. 3) Use (9) to obtain an estimate of each missing pixel em-
ploying the weights obtained in the previous step. 4) Iteratively
reestimate the missing pixels using (9) until convergence. Note
that since the cost function is convex, convergence is guaran-
teed.
For intra-coded blocks, the missing pixel values are initially
set to zero and then the MAP estimator, using the adaptiveMRF
and the data of the neighboring blocks, is applied. For missing
inter-coded blocks, an initial estimate are obtained using infor-
mation from the previous frame. Estimates of the prediction er-
rors are found using the adaptive MRF model along with the pre-
diction error signals of the neighboring blocks. The estimated
values are then added to the initial estimates. Since the predic-
tion error signal consists mostly of high-frequency components,
the MAP estimation stage (i.e., the second stage) will improve
the video reconstruction quality, especially around the edges.
IV. COMPUTATIONAL COMPLEXITY
The numbers provided below for computational complexity
correspond to a block and subblock of sizes 16 16 and 8
8, respectively. The computational load of the first stage of the
proposed method consists of those computations required in 1)
estimating the motion, and 2) computing the error of (6). For
motion estimation, we used a spiral search method using an area
of 16 16. This requires approximately 197 000 operations. If
the search for the displacement of each of the subblocks is re-
stricted to the set of candidate motion vectors explained in Sec-
tion III, the required number of operation will reduce to 6100.
The number of operations required to find the total error of (6)
for four blocks is approximately 380.
For the second stage, the computational load consists of
those computations required in adapting the weights of the
MRF model, which are approximately 18 400 operations for
a missing block, assuming four available neighboring blocks.
The estimation of missing pixels using (9) is an iterative
procedure which, for each iteration, requires approximately 30
operations for each missing pixel, or 16 16 30 operations
for each missing block. On average, 80 iterations are required
for the algorithm to converge for a block.
Clearly, the computational load of our error-concealment
method is quite reasonable. Our simulation experiments
confirm that the run time of our method is indeed acceptable.
In fact, real-time decoding (e.g., 10 frames/s for QCIF video
sequences) is still possible on a Pentium 300 MHz PC.
V. EXPERIMENTAL RESULTS
Although the method proposed in this paper is general and
can be applied to any block-based video compression method,
H.263 is used as our video coding framework. QCIF (176
144) video sequences at a temporal resolution of 10 frames/s are

1126 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 18, NO. 6, JUNE 2000
Fig. 3. PSNR values for image sequence Foreman with 10% GOB missing for
different concealment methods.
coded at 64 kbps and then decoded in the presence of slice/GOB
errors. The size of the blocks is 16 16, that of the subblock is
88.
To simulate the channel errors, the following tasks are
performed. Coded video information is first grouped into
packets, where each video packet consists of the coded data of
a GOB or a slice. The video packets are then multiplexed with
audio information according to the H.223 standard. We used
an H.223 multiplexing simulator which receives video packets,
simulates audio traffic, applies errors to the multiplexed bit
stream according to an error pattern stored in a file, and
outputs the packets [12]. The error pattern that we employed
corresponds to a mobile channel [13]. Burst errors will most
likely not corrupt two consecutive video packet, since audio
packets are inserted between them. The erroneous bit stream
is decoded such that the effect of errors will appear as missing
slices/GOB’s.
For inter-coded blocks, our error concealment consists of ob-
taining an initial estimate of the missing block using informa-
tion from the previous frame, and finding an estimate of the pre-
diction error using the adaptive MRF model. We compare the
performance of the above method with two other methods: 1)
a replacement method, where a missing block is replaced with
the same block in the previous frame; and 2) a median method,
where the motion vector of a missing block is set to the median
of motion vectors of blocks to the left, above and above-right
of it, and the estimated motion vector is used to obtain a mo-
tion compensation block which serves as the concealment of the
missing block. Here, we consider only the case where a GOB is
placed in each of the video packets. Shown in Fig. 3 are the con-
cealment PSNR values for the three methods mentioned above
for the video sequence Foreman with a 10% packet (GOB) loss
rate. As can be seen, for video sequences with a large amount of
motion like Foreman, the performance of the proposed method
and the median method are close to each other, and both are
better than the replacement method. Although the performance
Fig. 4. An inter-coded frame of the sequence Foreman concealed by the (a)
replacement, (b) median, and (c) proposed methods.
of the above methods is comparable in terms of PSNR, the
video sequences obtained using the proposed method are more
visually appealing than the ones obtained using the other two
methods. Fig. 4(a), (b), and (c) shows the reconstructed video
frame (inter-coded) of the video sequence Foreman. As can be
seen in the images, the replacement method does not generate
good results, especially around the nose area. Moreover, the
concealment result of the proposed method is clearly better than
that of the other two methods.
For intra-coded blocks, we compare the performance of our
method (which consists of estimating the missing block using
the adaptive MRF model) to that of four other efficientmethods:
1) a MAP estimator using a nonadaptive GMRF as the a priori
model where each missing pixel is basically set to the average
of the pixels around it, 2) a suboptimal version of the MAP esti-
mator where a missing pixel is set to the median of pixels around
it [8], 3) the deterministic method proposed in [4], and 4) the
deterministic method proposed in [3]. Although there are other
deterministic methods that are likely to outperform the above
two methods, for example, the ones proposed in [6] and [7], we
restrict our comparison to the above two deterministic methods.
In fact, a direct comparison of our method, which is statistical,
to deterministic methods is difficult because of different design
approaches. Our main aim here is to compare our method to pre-
viously proposed statistical methods.
In this experiment, we employ different video sequences and
different packetization schemes. In the first set of simulation
experiments, we assign a slice which consists of one block,
to a packet. Fig. 5(a) shows a frame of the image sequence
Foreman encoded and decoded using an H.263 compliant coder.
Fig. 5(b)–(g), shows (respectively) the same frame: 1) missing
approximately 20% of the packets (blocks), 2) reconstructed
using the nonadaptive GMRF model, 3) using the suboptimal
MRF model, 4) using the method proposed in [4], 5) using the
method proposed in [3], and 6) using our proposed adaptive
MRF error concealment method. Clearly, our proposed method
performs best in reconstruction quality, particularly in retrieving
the edges and in the areas of the frame that correspond to adja-
cent missing blocks.
In the next set of simulation experiments, we assign the coded
data of a GOB, which consists of 11 blocks, to each of the
video packets. Fig. 6(a) shows a frame of the video sequence
Foreman with two missing packets (GOB’s), at approximately
18% loss rate. Fig. 6(b) and (c) shows the error- concealment
result obtained using the nonadaptive GMRF model and sub-
optimal MRF models, respectively. Fig. 6(d) and (e) shows the
result obtained using the methods proposed in [4] and [3], re-
spectively. Fig. 6(f) shows the result obtained using our adaptive

SHIRANI et al.: CONCEALMENT METHOD FOR VIDEO COMMUNICATIONS 1127
Fig. 5. A frame from the video sequence Foreman: (a) original; (b) missing
blocks; reconstructed using (c) a nonadaptive GMRF model; (d) a suboptimal
MRF model; (e) the method proposed in [4]; (f) the method proposed in [3]; and
(g) our adaptive MRF model.
Fig. 6. A frame from the video sequence Foreman: (a) with missing GOB’s,
reconstructed using (b) a nonadaptive GMRF model; (c) a suboptimal MRF
model; (d) the method proposed in [4]; (e) the method proposed in [3]; and (f)
our adaptive MRF model.
MRF model. By comparing the figures, the superior perfor-
mance of our method becomes obvious. This performance ad-
vantage is demonstrated in the blocks that contain edges. For a
quantitative evaluation, Table I provides the PSNR values of the
above concealment methods (for both packetization cases) for
the video sequence Foreman. The table demonstrates that our
method outperforms the other methods by at least 2 dB.
Finally, we compare the computational complexity of our
method in terms of required number of additions and multiplica-
tions to the methods proposed in [4] and [3]. The total number
of additions and multiplications required for restoring a 16
16 block are approximately 4500 and 20 000 using the method
proposed in [4] and our method, respectively. The method pro-
posed in [3] requires approximately 650 000 additions and mul-
TABLE I
PSNR (dB) COMPARISON OF DIFFERENT
METHODS FOR THE VIDEO SEQUENCE FOREMAN
tiplications to restore a 16 16 block. Therefore, the number
of computations of our method is 4.4:1 larger than that of [4]
and 1:0.03 less than those of [3]. However, our method out-
performs [4] substantially in terms of the quality of the recon-
structed images. This can be seen in Figs. 5 and 6. This quality
improvement justifies, in most applications, the additional com-
putational complexity.
VI. CONCLUSION
This paper introduces a new method of error concealment
for block-based coded video communications over error-prone
channels. The proposed method employs an adaptive MRF
as the image a priori model in a MAP estimation paradigm.
The adaptation enables the estimation procedure to incorporate
more information without increasing the order of the MRF.
The proposed concealment method is able to restore missing
blocks located in smooth and low-frequency areas, as well as
in high-frequency and edge portions of a video frame. Our
concealment method achieves very good computation–per-
formance tradeoffs, and outperforms previously proposed
MRF-based concealment methods.
ACKNOWLEDGMENT
The authors would like to thank the anonymous reviewers for
their insightful comments and constructive suggestions. The au-
thors would also like to gratefully acknowledge the help and
support provided by G. Cote, their colleague in the Signal Pro-
cessing and Multimedia Group at the University of British Co-
lumbia.
REFERENCES
[1] M. Ghanbari and V. Seferidis, “Cell loss concealment in ATM video
codecs,” IEEE Trans.Circuit System Video Technol., vol. 3, pp. 238–247,
June 1993.
[2] W. M.Lam, A. R. Reibman, and B. Liu, “Recovery of lost or erroneously
received motion vectors,” in Proc. Int. Conf. Acoust., Speech, Signal
Processing, vol. V, 1993, pp. 417–420.
[3] Y. Wang, Q. F. Zhu, and L. Shaw, “Maximally smooth image recovery
in transform coding,” IEEE Trans. Commun., vol. 41, pp. 1544–1551,
Oct. 1993.
[4] S. S. Hemami and T. H. Y. Meng, “Transform coded image reconstruc-
tion exploiting interblock correlations,” IEEE Trans. Image Processing,
vol. 4, pp. 1023–1027, July 1995.
[5] H. Sun and W. Kwok, “Concealment of damaged block transform coded
images using projection onto convex sets,” IEEE Trans. Image Pro-
cessing, vol. 4, pp. 470–477, Apr. 1995.
[6] W. Kwok and H. Sun, “Multi-directional interpolation for spatial error
concealment,” IEEE Trans. Consumer Electron., vol. 39, pp. 455–460,
Aug. 1993.
[7] W. Zeng and B. Liu, “Geometric-structure-based error concealment with
novel applications in block-based low-bit-rate coding,” IEEE Trans. Cir-
cuits Syst. Video Technol., vol. 9, pp. 648–665, June 1999.

1128 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 18, NO. 6, JUNE 2000
[8] P. Salama, N. Shroff, E. J. Coyle, and E. J. Delp, “Error concealment
in encoded video streams,” in Signal Recovery Techniques for Image
and Video Compression and Transmission, N. P. Galatsanos and A. K.
Katsaggelos, Eds. Boston, MA: Kluwer Academic, 1998.
[9] R. Talluri, “Error resilient video coding in the MPEG-4 standard,” IEEE
Commun. Mag., vol. 26, pp. 112–119, June 1998.
[10] S. Geman and D. Geman, “Stochastic relaxation, Gibbs distribution, and
the Bayesian restoration of images,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 11, pp. 689–691, 1984.
[11] S. Shirani, F. Kossentini, and R. Ward, “Reconstruction of motion vector
missing macroblocks in H.263 encoded video transmission over lossy
networks,” in Proc. Int. Conf. Image Processing, vol. III, Oct. 1998, pp.
487–491.
[12] G. Sullivan, “A Simple Video Packet Mux Simulator Program for Video
Streams in H.263/M Using AL3 Mux of H.223 Annex B,” ITU-T Study
Group 15, Video Coding Expert Group, Doc. Q15-F-16, Nov. 1996.
[13] Ericsson Sweden, “WCDMA Error Patterns at 64 kb/s,” ITU-T Study
Group 16, Multimedia Terminals and Systems Expert Group, Cannes,
France, June 1998.
Shahram Shirani (S’88–M’89) received the B.Sc.
degree in electrical engineering from Esfahan
University of Technology, Esfahan, Iran, in 1989 and
the M.Sc. degree in biomedical engineering from
Amirkabir University of Technology, Tehran, Iran
in 1994. Currently, he is pursuing the Ph.D. degree
in the Department of Electrical and Computer Engi-
neering, University of British Columbia, Vancouver,
B.C., Canada.
From 1994 to 1996, he was with the Department
of Electrical Engineering, University of Tehran. His
research interests include image and video compression, video communications,
signal processing and ultrasonic imaging.
Faouzi Kossentini (S’89–M’89–SM’98) received
the B.S., M.S., and Ph.D. degrees from the Georgia
Institute of Technology, Atlanta, in 1989, 1990, and
1994, respectively.
During 1995, he was a Research Scientist at
Nichols Research Corporation, Huntsville, AL.
Since January 1996, he has been an Assistant
Professor and then an Associate Professor in the
Department of Electrical and Computer Engineering,
University of British Columbia, vancouver, B.C.,
Canada, where he is involved in research in the areas
of signal processing, communications, and multimedia. He has coauthored
more than 100 journal papers, conference papers, book chapters and patents.
Dr. Kossentini has been active as a Voting Member, and recently as head
of delegation, of the Canadian delegate to ISO/IEC JTC1/SC29, which is re-
sponsible for the standardization of coded representation of audiovisual, mul-
timedia, and hypermedia information. In particular, he has participated in most
current JBIG/JPEG and MPEG-4 standardization activities. He has also partic-
ipated in most current ITU-T low bit rate video coding standardization activi-
ties. He has served as a technical area coordinator and member of the technical
program committee of ICIP-1997, and as a Member of the technical program
committee of ISCAS-1999. He is also the Vice General Chairman of ICIP-2000.
Dr. Kossentini is currently an Associate Editor for the IEEE TRANSACTIONS ON
IMAGE PROCESSING. He is also an Associate Editor for the IEEE TRANSACTIONS
ON MULTIMEDIA.
Rabab Ward (F’98) received the B.Sc. degree in
electrical engineering from University of Cairo,
Egypt, in 1966 and the M.Sc. and the Ph.D. in elec-
trical and computer engineering from University of
California, Berkeley, in 1969 and 1972, respectively.
She is a Professor in the Electrical and Computer
Engineering Department, University of British Co-
lumbia, Vancouver, B.C., Canada, and is the Director
of the Center for Integrated Computer Systems Re-
search there. Her research interests are mainly in the
areas of signal processing and image processing. She
has made contributions in the areas of signal detection, image encoding, com-
pression, recognition, restoration and enhancement, and their applications to
infant cry signals, cable TV, HDTV, medical images, and astronomical images.
She holds five patents related to cable television picture monitoring, measure-
ment and noise reduction.
Dr. Ward is a fellow of the EIC and the Royal Society of Canada. She is
currently serving as the General Chair of ICIP’2000 to be held in Vancouver.