Science Service System

Summary of Proposal MTH0036

TitleCompression of high resolution SAR data
Investigator Gleich, Dušan - University of Maribor, Faculty of Electrical Engineering and Computer Science, Laboratory for Signal processing and remote control
Team Members
dr. Cucej, Zarko - University of Maribor, Faculty of Electrical Engineering and Computer Science
dr. Gleich, Dušan - University of Maribor, Faculty of Electricel Engineering and Computer Science
Mr. Hebar, Marko - University of MAribor, Faculty of Electrical Engineering and Computer Science
dr. Kseneman, Matej - University of Maribor, Faculty of Electrical Engineering and Computer Science
SummaryThe project objective is to compress high resolution SAR data and extraction of data features. It is well known that for high compression efficiency and ''good'' feature extraction data models must be well specified. The compression codec must be able to eliminate data redundancy meanwhile for feature extraction is to find a model that will best describe useful data. Therefore our goal is to join data modeling for feature extraction with method for redundancy elimination and compression method. There are many methods for data compression and feature extraction. One of them is vector quantization that has been widely used in SAR image processing community. The problem of vector quantization is that it is very high computationally demanding. However, there have been many attempts to speed up a vector quantization. Our goal is to compress and extract features in the wavelet domain. The wavelet domain is interesting signal processing tool since it is able to spatially de-correlate data. The property of wavelet transform is that the experimental probability density function (pdf) of subbands has general Gaussian distribution, but the original data has usually gamma or K distribution. The wavelet subbands can be easily modeled using theoretical general Gauss distribution. Here must consider that SAR data are corrupted by a multiplicative noise call spackles. Image and noise models must be defined here. We have shown that noise can be efficiently removed from SAR image in wavelet domain using Gauss Markov Random fields and Bayesian inference. The result of our research is also that speckle noise can be efficiently removed using mixture models of Gaussians. The theory of powerful Monte Carlo methods will be applied here in order to find models for SAR image de-noising and features extraction. The models will have to be able to deal with complex and real data. Obtained models will also be used for transform coding of wavelet coefficients. Transform coding consists of transformation, quantization and entropy coding. The quantization techniques will be researched that preserved the wavelet coefficients in order to preserve extracted features. Models used here will be fused with quantization models such as trellis quantization and context based quantization. We will use our previous knowledge to developed context based entropy coder that will be able to efficiently code quantized indices. As we mentioned, we will compress complex single look SAR data, intensity and phase part SAR data. Our goal is to estimate how the compression impacts on further SAR analysis and what is the compression ratio can be achieved and how the compression impacts on extracted data. The work will be financed by Ministry of higher Education and Science, Slovenia.

Back to list of proposals

© DLR 2004-2016