| Summary: | 碩士 === 國立成功大學 === 電機工程研究所 === 82 === In this thesis, a new pratical technique for image data
compression in vector quantization codebook design is
presented. Vector quantization (VQ) design is a
multidimensional optimization problem in which the
specification of the codebook plays a major role in deciding
the quality of reconstructed images. The traditional method for
VQ codebook design is the Generalized Lloyd Algorithm (GLA), an
iterative descent algorithm which monotonically decreases the
distortion function towards a local minimum, so we develop a
new algorithm for this. A procedure named Simulated Annealing
(SA) is a technique for solving combinatorial optimization
problems, so this approach should lead to a good solution. In
other words, using SA to design codebook can find a global
minimum; however, the amount of computation required is
extremely large since many switches must be attempted at each
temperature. We will discuss how to reduce the amount of
computation needed by some strategies. In fact, even these
approaches still take a very long time to converge since in
large problems switching one codevector has quite a small
effect on the total distortion and convergence. Finally, we
consider an "implementable" approach named SA-GLA by using GLA
to get a local minimum at each temperature firstly then to
judge whether it reach stable or not by above SA until stopping
at some minimum energy state. Experimental results are
presented in terms of bit rates and the qualities of the
reconstructed images. It can be shown that the SA-GLA is a
reliable way to improve yhe quality of the codebooks used for
VQ, and the improvement obtained with this algorithm over the
GLA inevitably varies with the initial codebook. Comparing SA-
GLA with GLA, under bit rate = 0.5 bpp, for a natural image
LENA the SNR value we can improve is 0.6 dB.
|
|---|
