Databases Reference
In-Depth Information
Stage 1
Stage 2
Stage S
c
c
2,1
c
S,1
1,1
residual
c
1,2
c
1,3
residual
2
residual
c
2,2
c
2,3
c
S,2
c
S,3
residual
LSF
S−1
1
c
1,M1
c
2,M2
c
S,MS
F I GU R E 18 . 13
Multistage vector quantizer for the quantization of LSF coefficients
in the SILK encoder.
of noise power occupies the same region as the highest amount of signal power. Thus even if
the total noise power is not reduced the perceptual effect of the noise can be greatly reduced.
The LSF coefficients are obtained in a manner similar to the other methods discussed
earlier. However, the quantization process is significantly more complex. The vector of LSF
coefficients is quantized using a multistage quantizer as shown in Figure
18.13
. Each stage
of the quantizer operates on the residual of the previous stage. Thus the input of the second
stage is the difference between the LSF coefficients and the code vectors of the first
M
1level
vector quantizer. A greedy approach to multistage vector quantization is to simply take the best
representative at each stage. However, this reduces the efficacy of the quantization process.
Ideally, we would like to keep the residual from each stage until the final stage at which time we
could make a selection among all possible combinations. However, this would mean keeping
track of an exponentially growing number of quantized representations. The total number of
possible combinations in the SILK algorithm for sampling rates above 8 kHz is 2
36
! The SILK
algorithm allows this to be a dynamic process where at each stage the most representative
code vectors, called the survivors, are allowed to proceed to the next stage. At each stage the
weighted sum of the accumulated bit rate and accumulated distortion for each code vector is
evaluated to determine the survivors for the next stage. The number of survivors is adjusted
based on available resources.
The SILK encoder uses a variable-rate entropy coder to encode the various coding param-
eters. The range coder is an analog of the arithmetic coder that uses a nonbinary alphabet.
This helps improve the rate distortion performance of the coder while significantly increasing
the complexity of the coder.
18.6 Image Compression
Although there have been a number of attempts to mimic the linear predictive coding approach
for image compression, they have not been overly successful. A major reason for this is that
while speech can be modeled as the output of a linear filter, most images cannot. However,
a totally different analysis/synthesis approach, conceived in the mid-1980s, has found some
degree of success—fractal compression.
