22-06-2012, 05:59 PM
Loop-filtering and Post-filtering for Low Bit-rates Moving Picture Coding
[attachment=25158]
Introduction
Most video compression standards, including ITU-T
H.263 [ 1-21 and MPEG-4 [3], use a block-based discrete
cosine transform (DCT) and a block-based motion
compensation (MC). The 8x8 block-based DCT scheme,
which is used to pack information into a few transform
coefficients, takes advantage of the spatial correlation
property of images. However, this block-based coding
induces the well-known blocking artifacts, comer
outliers, and ringing noise, particularly, when the image
is highly compressed. The blocking artifacts are the grid
noise along the block boundaries in a relatively
homogeneous area. Comer outliers occur at the comer
points of the 8x8 blocks, and the ringing noise is due to
truncation of the high-frequency coefficients by
quantization.
The proposed filtering methods
Deblocking filter
One-D LPF to reduce the blocking artifacts is
performed strongly or weakly, depending on the blocking
flags, on the horizontal block boundary and on the
vertical block boundary. The proposed deblocking
algorithm does not require image-edge detection, which
requires a large number of computations, because it
utilizes the block-based blocking flags obtained in
Section 2.
Compensation of corner outliers
Corner outlier compensation is only performed in
INTRA frame. A corner outlier [7] is characterized by a
pixel which is either much larger or much smaller than its
neighboring pixels in the comer point of an 8x8 block of
the decompressed image. When a dark-gray region is
distributed over four blocks and one or two pixels of the
dark-gray region are located in the corner points of
neighboring blocks, the corner points can be distorted by
quantization of the DCT coefficients. This point is called
a comer outlier. In order to reduce the comer outlier, the
corner outlier must be detected and compensated for [7].
Simulation results
The simulation was performed with unrestricted
motion estimation, advanced prediction mode, advanced
INTRA coding mode and a TMN8 rate control of H.263+
TMN10 (H.263 version 2) [3]. QCIF image sequences,
each of which had 300 frames, were used for this
simulation.
[attachment=25158]
Introduction
Most video compression standards, including ITU-T
H.263 [ 1-21 and MPEG-4 [3], use a block-based discrete
cosine transform (DCT) and a block-based motion
compensation (MC). The 8x8 block-based DCT scheme,
which is used to pack information into a few transform
coefficients, takes advantage of the spatial correlation
property of images. However, this block-based coding
induces the well-known blocking artifacts, comer
outliers, and ringing noise, particularly, when the image
is highly compressed. The blocking artifacts are the grid
noise along the block boundaries in a relatively
homogeneous area. Comer outliers occur at the comer
points of the 8x8 blocks, and the ringing noise is due to
truncation of the high-frequency coefficients by
quantization.
The proposed filtering methods
Deblocking filter
One-D LPF to reduce the blocking artifacts is
performed strongly or weakly, depending on the blocking
flags, on the horizontal block boundary and on the
vertical block boundary. The proposed deblocking
algorithm does not require image-edge detection, which
requires a large number of computations, because it
utilizes the block-based blocking flags obtained in
Section 2.
Compensation of corner outliers
Corner outlier compensation is only performed in
INTRA frame. A corner outlier [7] is characterized by a
pixel which is either much larger or much smaller than its
neighboring pixels in the comer point of an 8x8 block of
the decompressed image. When a dark-gray region is
distributed over four blocks and one or two pixels of the
dark-gray region are located in the corner points of
neighboring blocks, the corner points can be distorted by
quantization of the DCT coefficients. This point is called
a comer outlier. In order to reduce the comer outlier, the
corner outlier must be detected and compensated for [7].
Simulation results
The simulation was performed with unrestricted
motion estimation, advanced prediction mode, advanced
INTRA coding mode and a TMN8 rate control of H.263+
TMN10 (H.263 version 2) [3]. QCIF image sequences,
each of which had 300 frames, were used for this
simulation.