SIGNAL ADAPTIVE POSTPROCESSING FOR BLOCKING EFFECTS
REDUCTION IN JPEG IMAGE
H.
C. Kim* and
H.
W.
Park
Dept. Informatiom and Communication Eng., KAIST
207-43 Cheongryangni, Dongdaemungu
Seoul 130-012, Korea
hwparkasicom. kaist.ac.kr
*LG
Information
&
Communication, Ltd.
ABSTRACT
A
postprocessing algorithm is proposed to reduce the
blocking artifacts of joint photographic experts group
(PEG) decompressed images. The reconstructed images
from
JPEG
compression produce noticeable image
degradation near the block boundaries, in particular, for
highly compressed images because each block is
transformed and quantized independently. The reduction
of these blocking effects has been an essential issue for
high quality visual communications. The proposed
postprocessing algorithm reduces these blocking artifacts
efficiently.
A
comparison study between the proposed
algorithm and other postprocessing algorithms
is
made
by computer simulation with several PEG images. These
simulation results show that the proposed algorithm
reduces the blocking artifacts significantly in the
subjective and objective views.
I.
INTRODUCTION
PEG
had been recommended as a standard
compression scheme for continuous-tone still images
[
11.
JPEG
uses an 8x8 pixel-block discrete cosine transform
(DCT) for information paclung into a few transform
coefficients. This block DCT scheme takes advantage of
the local spatial correlation property of images and also
saves processing time [2]. However,
it
is well
known
that
this individual processing of each block induces visually
annoying blocking effects [3][4][5], in particular, when a
large quantization parameter is used for high
compression.
There are three kinds of blocking effects in
PEG
decompressed images. One is the staircase noise along
the image edges, another is the grid noise in the
monotone area, and the other is the corner outlier in the
corner point
of
the 8x8 DCT block. When an 8x8 block
includes an image edge, the edge is degraded such that
the block boundary looks like the edge. This artifact is
0-7803-3258-X/96/$5.00
0
1996 IEEE
41
called staircase noise. A slight change
of
image intensity
between the 8x8 block boundaries is easily noticeable in
the monotone area. This change is called grid noise. A
corner outlier is visible at the corner point of the 8x8
block, where the corner point is either much larger or
much smaller than the neighboring pixels.
Several postprocessing algorithms have been proposed
to reduce the blocking effects
of
block-based coding
[3] [4].
A
linear space-invariant low-pass filtering can
reduce the blocking effects, but
it
also degrades image
details such as edge and texture, and reduces the image
contrast.
In
order to overcome this degradation problem,
a signal adaptive filter can
be
applied to the
PEG
decompressed image. In the signal adaptive filter, the
filter frequency response should
be
varied in accordance
with the local signal and noise characteristics.
In
the
Ramamwthi algorithm[3], the ratio of the intensity
change
is
compared to the average intensity with a
threshold.
In this paper, a new postprocessing algorithm based
on a signal adaptive filtering and
on
a corner outlier
detectionheplacement scheme is proposed. In order to
preserve the image details, a local threshold value is
defined by using the local variance, mean, and global
threshold of the grachent image [6][7][8][9]. The
staircase noise is smoothed
by
a 1-D drectional filter
accordmg to the directions of the edge. The grid noise is
smoothed by a signal adaptive
2-D
low-pass filtering
whose kernel size is 5x5. The comer outlier [lO][ll]
detectionheplacement scheme consists of two stages. At
the first stage, the corner outlier is detected by using the
intensity difference between the corner point of the 8x8
block and its neighbors. At the second stage, the detected
corner outlier and its neighbors are replaced with new
values calculated by the weighted average method.
Section
2
describes the proposed postprocessing
algorithm for blocking effect reduction in
JPEG
decompressed images. Computer simulations was
performed to verify the performance of the proposed
algorithm, whose results are presented in section 3.
Finally, conclusions are given in section 4.
Sobel
operation
JPEG
decompressed
image
*
Gradient
image
Global
Classification
-
Local
edge
map
1
-contourmap
Threshold
I
J
Postprocessed
Image
Fig.
1.
Overall block diagram
of
the proposed postprocessing
algorithm.
OPOSED ALGORITHM FOR
OF
BLOCKING EFFECTS
2.1.
The Property of the
Gradient
EG Decompressed Image
Figure
1
shows the overall block dagram of the
proposed algorithm. The proposed algorithm is based on
the edge information, which is the thresholded gradent
image. Basically, the idea underlying most edge detection
methods is the computation of a derivative, i.e., image
hfferentiation. Gradent operators measure the gradient
of the image in a specified direction [6]
[7].
There is an
interesting property of the gradient absolute image
computed by the Sobel gradient operator, which is shown
in Fig. 2. The gray level of the grahent absolute image
has the property that R=S=2Q, where R is a gray level
in the gradient absolute image,
S
is the slot interval in
the histogram of Fig.'2.a, and Q is the quantization factor
I
4R
I
4R
I
1
4R
I
4R
I
Fig
2
Property
of
the
PEG
decompressed mage (a) Histogram
of
the
PEG
decompressed image,(b) Magdied view
of
the
grid noise property
of
the gradient absolute image where each
square is a pixel
in the
JPEG
compression.
As
shown
in
Fig. 2.b, all grid
noises in the monotone area are less than 6R. For
example, when the
Q
factor is set to
8,
all grid noises in
the gradent image are less than
100.
Therefore, if we
choose a global threshold value of
100,
the grid noise
will not be included in the global edge map.
2.2.
Classification
The
JPEG
decompressed image is classified into two
areas, such as an edge area and a monotone area,
accordmg to the edge map, which
is generated by
thresholding the gradient absolute image. In general,
thresholding is a typical approach to image classification.
A
global threshold value
Tg
,
which has a value of 100 in
this paper, for the global edge map is chosen in
consideration with the
Q
factor and the histogram
characteristics
of
the gradient image. The performance of
the proposed algorithm does not critically depend on the
global threshold value.
In addtion to the global threshold, local thresholding
is performed for the local edge map of each
8x8
block.
The local threshold value
T,
of
the n-th
8x8
block is
defined by the characteristics of the pixels within the
block It can be calculated as follows
[7],
The parameters
m,
and onare the mean and the standard
deviation of the n-th 8x8 block of the gradent absolute
image, respectively.
T,
is the global threshold value.
If the n-th DCT block
is
homogeneous, the ratio
ma
ink
tends to be zero with the result that the threshold value
T,
tends toward
Tg.
If the n-th DCT block is complex, the
ratioq
fwk
is increased, resulting in the threshold value
T,
being smaller than
Tg.
This small T, generates a detailed
edge map which is not classified as a global edge by
T,.
Therefore. the local edge map is computed by locally
thresholdmg the 6x6 pixels within the 8x8 block
excludmg boundary pixels in order to preserve the
detailed information from blurring and to prevent the
grid noise of the block boundary from being detected as
an
image edge.
2.3.
Signal
Adaptive Filter
A
signal
adaptive filter is proposed to reduce both the
grid noise in the relatively monotone area and the
staircase noise along the image edge without any
signgicant loss of the image details.
In
the first step of
42
the signal adaptive filtering, a structure window scans the
given global edge map in order to obtain the contour
map as shown in Fig.
3.
If
the origin and the other four
points in the structure window are all included in the
global edge map, the origin point is excluded from the
contour map. After the contour map is obtained from the
global edge map, 1-D directional filtering of the PEG
decompressed image is performed for all points on the
contour map in order to reduce the staircase noise.
A
1-D directional smoothing filter aligned parallel to
the edge is ideally suited for the reduction of staircase
noise, which is visible along the image edge. The three
coefficients of the 1-D smoothing filter are 1,4, and 1.
If,+
Origin
Fig.
3.
Contour pixel classification scheme for one-dimensional
filtering: (a) Structure window,
(b)
Global edge map;
(c)
Contour map.
In
the second step, signal adaptive
2-D
low-pass
filtering
is
performed. When the central point, point
11,
of the filter kernel in Fig. 4.a is
on
the global or the local
edge map,
no
2-D filtering
is
performed for this point. If
any edge point of the global or the local edge map is not
included in the 5x5 filter window, unweighted average
filtering is performed
by
using kernel 1 in Fig. 4.b. If
some edge pixels are in the 5x5 filter window except
on
the central point, the weighted average filtering is
performed by using kernel 2 in Fig. 4.c.
In
the filter
operation using kernel
2,
if any edge point is
on
the
points 12, 7, 6, 5, 10. 15, 16, or 17 in Fig. 4.a, the
weights of the edge pixel and its outer neighbor pixels
are set to zero. The weight value is changed according to
the position of the edge pixel in the filter window to
protect the image details.
Fig.4. Signal adaptive
2-D
filter kernels: (a) The coordination
of a signal adaptive
2-D
LPF
kernel;
(b)
Kernel
1;
(c) Kernel
2.
2.4.
Comer Outlier Detection and Replacement
A
corner outlier [lO][ll] is characterized by a pixel
which is either much larger or much smaller than its
neighboring pixels
in
the corner point
of
the 8x8 DCT
block
in
the PEG decompressed image. Each pixel in the
corner point is compared with its neighbors in order to
detect the corner outlier.
A detected corner outlier and
adjacent pixel will
be
replaced by the weighted average of
adjacent corner points with the ratio
of
3:l:l:l. This
detection and replacement process is performed in the
JPEG decompressed image.
3.
SIMULATION RESULTS
The proposed postprocessing algorithm was applied to
sample images such as Lena, Pepper, and Zelda, all
of
which were compressed with the standard JPEG
algorithm. Resolutions
of
the images were all 5 12x5 12
pixels with 8-bit gray-levels. Graphical and comparison
results are presented for only the Lena image in this
paper, since the simulation results for the other images
are similar to those of the Lena image.
PSNR
of the postprocessed image from the
Ramamurthi algorithm and the proposed algorithm are
summarized in Fig.
5.
The performance of the proposed
algorithm
is
better than that of the Ramamurthi
algorithm. The gradient absolute images obtained by the
Sobel operation can show the performance of the grid
noise reduction by the postprocessing algorithms. The
grid noise along the 8x8 DCT block boundary is most
visible in the PEG decompressed image, whereas the
noise is mostly removed
in
the postprocessed image from
the proposed algorithm as shown in Fig.6.
I
PSNR
versus bit per pixel curve
of
the Lena
32
31
30
E.
z
29
28
27
2
26
-0-
Proposed
-A-
Ramamurthl
..
.
0.335525 0,295849 0.268283
0.497634 0.393711
Q=Z
Q=3
Q=4
Q-5
Q=6
Bit
per
pixel and quantization factor
Fig.5.
PSNR
comparison between the PEG decompressed
image, the postprocessed image
by
the
Ramamurthi
algorithm,
and the postprocessed image
by the proposed algorithm for the
Lena image.
43
Fig
6
The gradient absolute image
of
the
Lena
mage
(a)
Gradient
absolute
image
of
the
PEG
decompressed
image,
lena
jpg, (b) Gradient absolute image
of
the postprocessed
image
by
the
proposed
algorithm
4.
CONCLUSIONS
In this paper, a new blocking-effect-reduction
algorithm was proposed to improve the quality of the
PEG decompressed image.
To
reduce the blocking
effects without degradation
of
the image details, the
proposed algorithm uses signal adaptive filtering
as well
as a corner outlier detection and replacement scheme.
The objective performance of the proposed algorithm
was measured by the PSNR of the postprocessed image
from the proposed algorithm. The measured PSNR
showed an increase of 0.2
-
1.0
dB
for the three test
images of the Lena, the Pepper and the Zelda images at
several compression ratios. However, the PSNR does not
fully reflect the enhancement
of
the image quality in the
psychovisual viewpoint. From a subjective view. the
blocking effects were almost removed in the gradient
absolute image made
by
a Sobel operation. In addtion, a
visual improvement
of
the image quality was found in the
postprocessed image from the proposed algorithm.
In conclusion, the proposed postprocessing algorithm
effectively reduces the blocking effects and well preserves
and enhances
PEG
images without any increase in the
bit rates.
5.
REFERENCES
El]
W.
B.
Pennebaker and J. L. Mitchell, PEG Still
Image Data Compression Standard, Van Nostrand
Reinhold, New York: 1993.
121
M. A. Sid-Ahmed, Image Processing, McGRAW-
HILL,
New York, 1994.
[3] B. Ramamurthi and A. Gersho, “Nonlinear Space
Variant Postprocessing of Block Coded Images,”
IEEE
Trans. on ASSP, vol. 34, no.
5,
pp.
1258-
1267, 1986.
[4]
B. Zeng and A. N. Venetsanopoulos, “A PEG-
based Interpolative Image Coding Scheme,”
[5]
W.
E. Lynch, A. R.
Reibman, and
B. Liu, “Post
Processing Transform Coded Images using Edges,’‘
[61
A.
K.
Jain, Fundamentals of Digital Image
Processing, Prentice-Hall, New Jersey, 1989.
[7]
I. Pitas, Digital Image Processing Algorithms,
Prentice-Hall, New York, 1993.
[8]
W. Lynch, A. R .Reibman, and B. Liu, “Edge
Compensated Transform Codng,” ICASSP,
pp.
191
S.
M. Kay and G.
J.
Lemay, “Edge Detection Using
The Linear Model,” IEEE Trans. on ASSP,
vol.
ASSP-34, no.
5,
pp.
1221-1227, 1986.
[lo] W.
K.
Pratt, Digital Image Processing, John Wiley
&
Sons,
New
York,
1991.
[ll]
X.
You and G. Crebbin,
“A
Robust Adaptive
Estimator for Filtering Noise in Images,” IEEE
Trans. on Image Processing
..
vol. 4, no. 5,
pp.
693-
699, 1995.
ICASSP,
pp.
393-396, 1993,
ICASSP,
pp.
2323-2326, 1995.
105-109, 1994.
44