| Transmission of large images, such as weather images obtained from satellites, to smaller ocean-going vessels is still expensive due to the rates charged for available communications services. Thus, low bit rate compression is very attractive to this market. In this dissertation, we explore the use of entropy-constrained vector quantization (ECVQ) to best represent the regions of interest in these satellite images at low bit rates. We also develop fast full search methods for vector quantizers using Lagrangian distortion measures, including ECVQ and Bayes-risk VQ.;ECVQ trades off distortion and rate using a Lagrangian formulation as its distortion measure. An ECVQ codebook depends on a good match with the statistics of the encoded image. A real world image can be modeled as a collection of regions whose statistics are stationary within the region but differ between regions. Our images are grayscale satellite images with slowly varying regions representing clouds, land, and water. The regions are, in this simple case, largely separable by their intensity. We examine the use of a combination of ECVQ codebooks to achieve the best match for each region. |
