Digital Signal Processing Reference
In-Depth Information
(a)
(b)
FIGURE 3.9
(a) ISAR image of B-727 and (b) portion of synthetic original and noisy training images used for
optimization.
to adjacent blocks, while the ringing artifacts are attributable to Gibbs phe-
nomenon and caused by the lossy quantization of frequency coefficients.
Many spatial and transform domain deblocking techniques have been
developed.
12
However, most techniques either (1) have high computational
complexity or (2) are limited to addressing certain types of artifacts. It is im-
portant to note that blocking artifacts manifest as relatively weak noisy edges.
The center affine filter can thus be used to smooth these artifacts while simul-
taneously sharpening the stronger true edges. Additionally, the complexity
of the center affine filter is only slightly greater than that of the linear filter
and significantly less than most deblocking algorithms.
To evaluate the ability of the center affine filter to reduce blocking artifacts,
we apply the method to JPEG images with various quality factors (QF). The
QF maps to a scaling parameter of the standard quantization table to control
the compression ratio. The range of QF is from 1 to 100, where 100 indicates no
quantization. Note that different compression ratios lead to different noise,
or artifact, levels. The center affine filter membership function and spread
parameter can be optimized for a specified compression level. Here we use a
Gaussian membership function and uniform spatial weights for all cases, but
allow the spread parameter to vary with the compression level.
We compare the center affine method with that proposed in Reference 13,
which we denote as the “2Mod” method. This provides a fair comparison as
both methods have similar computational complexity and the 2Mod method
is adopted in the MPEG-4 standard. The size of all test images is 256
256. For
each QF, we use Lenna as the training image to optimize the spread parameter
×
Search WWH ::
Custom Search