Databases Reference
In-Depth Information
1.
“Vector Quantization,” by R.M. Gray, in the April 1984 issue of
IEEE Acoustics, Speech,
and Signal Processing Magazine
[
273
].
2.
“Vector Quantization: A Pattern Matching Technique for Speech Coding,” by A. Gersho
and V. Cuperman, in the December 1983 issue of
IEEE Communications Magazine
[
162
].
3.
“Vector Quantization in Speech Coding,” by J. Makhoul, S. Roucos, and H. Gish, in the
November 1985 issue of the
Proceedings of the IEEE
[
167
].
4.
“Vector Quantization,” by P.F. Swaszek, in
Communications and Networks
, edited by
I.F. Blake and H.V. Poor [
274
].
5.
A survey of various image-coding applications of vector quantization can be found
in “Image Coding Using Vector Quantization: A Review,” by N.M. Nasrabadi and
R.A. King, in the August 1988 issue of the
IEEE Transactions on Communications
[
168
].
6.
A thorough review of lattice vector quantization can be found in “Lattice Quantization,”
by J.D. Gibson and K. Sayood, in
Advances in Electronics and Electron Physics
[
152
].
The area of vector quantization is an active one, and new techniques that use vector quanti-
zation are continually being developed. The journals that report work in this area include
IEEE
Transactions on Information Theory
,
IEEE Transactions on Communications
,
IEEE Transac-
tions on Signal Processing
, and
IEEE Transactions on Image Processing
, among others.
10.10 Projects and Problems
1.
In Example
10.3.2
we increased the SNR by about 0.3 dB by moving the top-left output
point to the origin. What would happen if we moved the output points at the four corners
to the positions
? As in the example, assume the input has a Laplacian
distribution with mean zero and variance one, and
(
±
,
0
), (
0
,
±
)
=
0
.
7309. You can obtain the
answer analytically or through simulation.
2.
For the quantizer of the previous problem, rather thanmoving the output points to
(
±
,
0
)
(
,
±
)
, we could have moved them to other positions that might have provided a
larger increase in SNR. Write a program to test different (reasonable) possibilities and
report on the best and worst cases.
3.
In the program
trainvq.c
the empty cell problem is resolved by replacing the vector
with no associated training set vectors with a training set vector from the quantization
region with the largest number of vectors.
and
0
In this problem, we will investigate some
possible alternatives.
Generate a sequence of pseudorandom numbers with a triangular distribution between 0
and 2. (You can obtain a random number with a triangular distribution by adding two
uniformly distributed random numbers.) Design an eight-level, two-dimensional vector
quantizer with the initial codebook shown in Table
10.9
.
(a)
Use the
trainvq
program to generate a codebook with 10,000 random numbers
as the training set. Comment on the final codebook you obtain. Plot the elements of
the codebook and discuss why they ended up where they did.
