Design of cubic lattice vector quantizers /
| Main Author: | |
|---|---|
| Other Authors: | , , |
| Format: | Thesis Book |
| Language: | English |
| Published: |
1990.
|
| Subjects: | |
| Online Access: | ProQuest, Abstract Link to OAKTrust copy |
| Abstract: | A cubic lattice is introduced in the design of n-dimensional vector quantizers (VQ's). Its simple structure results in structured VQ codebooks which allow for fast vector quantization of the input vectors by round up or down operations. The geometrical properties of the pdf's for both memoryless Gaussian and Laplacian vectors are examined and exploited in the VQ design. The resulting VQ for the Gaussian (Laplacian) source is called a spherical (pyramidal) uniform cubic lattice (Z^n) VQ. Simulation results show that, for both sources, the multidimensional cubic lattice VQ's outperform the optimum uniform scalar quantizer. It is also shown that, for the Laplacian source, a pyramidal codebook becomes far superior to a spherical codebook as the dimension and rate are increased. In order to reduce an excessive overload distortion caused by a constant scale factor in the uniform cubic lattice (Z^n) VQ's, quasi-uniform cubic lattice VQ's are constructed, where codebooks are formed by concatenating a series of uniform cubic lattice subcodebooks such that the respective nearest neighbor distance increases by a multiplication factor. This approach provides a structured codebook with both an affordable complexity and a performance improvement over the uniform cubic lattice VQ for both the Gaussian and Laplacian sources. Compared to the rate distortion performance, the quasi-uniform Z^16 VQ performs 2.26 and 2.00 dB below the rate distortion function for the Laplacian and Gaussian sources, respectively. The simulation results show that the quasi-uniform companding method is efficient for a source like the Laplacian source whose pdf has a long and heavy tail. The uniform Z^16 VQ's and the quasi-uniform Z^16 VQ's are applied to the encoding of the DCT coefficients of several 256 x 256 monochrome images at 0.5 bits per pixel. Both cubic lattice VQ's show significant performance improvements over scalar quantizers, and are comparable to an LBG VQ designed on a set of 10 images. Both cubic lattice VQ's are shown to be effective in image transform coding at low rates. |
|---|---|
| Item Description: | Typescript (photocopy). Vita. "Major subject: Electrical engineering." |
| Physical Description: | xii, 133 leaves : illustrations ; 29 cm |
| Bibliography: | Includes bibliographical references. |