Projects / Programmes
Complex analysis of high resolution SAR images
| Code |
Science |
Field |
Subfield |
| 2.15.02 |
Engineering sciences and technologies |
Metrology |
Signals and transmission |
| Code |
Science |
Field |
| T121 |
Technological sciences |
Signal processing |
Signal Processing, Bayes theory, Classification, Sequential Monte Carlo methods, Complex analysis
Researchers (6)
| no. |
Code |
Name and surname |
Research area |
Role |
Period |
No. of publicationsNo. of publications |
| 1. |
27565 |
PhD Karl Benkič |
Computer science and informatics |
Junior researcher |
2007 - 2008 |
0 |
| 2. |
20862 |
PhD Dušan Gleich |
Systems and cybernetics |
Head |
2007 - 2008 |
0 |
| 3. |
27986 |
PhD Marko Hebar |
Systems and cybernetics |
Researcher |
2007 - 2008 |
0 |
| 4. |
11576 |
PhD Jože Mohorko |
Telecommunications |
Researcher |
2007 - 2008 |
0 |
| 5. |
04799 |
PhD Peter Planinšič |
Systems and cybernetics |
Researcher |
2007 - 2008 |
0 |
| 6. |
19508 |
MSc Tomaž Romih |
Computer science and informatics |
Researcher |
2007 - 2008 |
0 |
Organisations (1)
Abstract
Statistical modeling of SAR scene is proposed in this research project. The SAR scene modeling, classification and denoising will be implemented in complex-wavelet domain using complex models and complex computations. Our goal is to denoise high resolution SAR images whilst preserving all SAR image characteristics as textures, edges, etc. The complex wavelet transform will be used on a complex SAR image and on intensity part of SAR image. The Bayesian inference will be used for denoising, classification. In order to denoise SAR image a prior and likelihood probability density functions (pdf) must be defined. SAR image contains a multiplicative noise called speckle. For a prior a generalized Gauss-Markov Fields will be used and for a noise model a General Gaussian pdf will be chosen. The noise model is usually multiplicative, but in our research we will assume that noise is additive and signal depended.
The noise free wavelet coefficient will be estimated by finding a maximum a posteriori (MAP) estimate. A second order Bayesian inference will be used to find the best model among all models. The model parameters will be changed using an Expectation Maximization algorithm and the maximization of the evidence. The texture parameter is presented by parameters of GGMRF. An algorithm for unsupervised Bayesian learning will be developed in order to recognize image areas with pre-determined textural parameters.
In the last part of the research project we will use the particle filters for denoising and classification of SAR images. Particle filters are sequential Monte Carlo methods based on point mass representation of probability density function. Our goal will be sampling particles from a prior distribution and assigning weights according likelihood pdf. Different sequential sampling methods will be implemented in this research project and the research results will show which method is the most appropriate. The algorithm for estimating textural parameters with particle filters is proposed in this project. The parameters of GGMRF will be changed sequentially in order to produce the highest number of particles.
Significance for science
In this research project we have developed methods for despeckling od Synthetic aperture Radar (SAR) images. ) SAR imaging is very popular because it provides images of the Earth’s surface in all weather conditions, day and night. Nowadays SAR images obtained from satellite radar platforms have resolution below 1 meter. SAR images contain multiplicative noise, which is called speckle noise. Speckles occur by random interferences, either constructive or destructive, between electromagnetic waves from different reflections in the imaged area. Multiplicative noise in SAR images makes interpretation and scene analyses very difficult; therefore, the goal of despeckling is to remove noise and to preserve all textural features in the SAR images. SAR images are corrupted by a multiplicative noise called 'speckle'. Speckle noise arises from image formation under coherent radiation. The presence of speckle noise in SAR images is undesirable since it makes scene analysis and understanding very difficult. The goal of despeckling algorithms is to remove speckle-noise whilst preserving all the image's textural features. Quantitative evaluation of reconstructed-denoised image includes several criteria such as equivalent number of looks, mean bias, edge and texture preservation, and computation complexity. Many different techniques for speckle removing or the denoising of SAR images have been proposed over the past few years. The denoising of SAR images can be performed in original-image domain or transformed-wavelet domain.
The research project has proposed methods for despeckling of complex SAR images within the complex wavelet and complex spatial domains. The product od despeckling is a complex image, which distribution corresponds to the model, which is used for despeckling. The model-based approach transforms the distribution of SAR images form Gamma distribution to the distribution which corresponds to the distribution of the prior model. This enables that despeckled images can be used together with its textural parameters for supervised classification and enables automatic scene interpretation.
Significance for the country
This project enables automatic scene interpretation. The developed methods in this research project enables automatic scene interpretation using texture models. The despeckled images helps to understand the scene more efficiently and to have a prior knowledge about the Earth's surface. The texture parameters of the texture-based models enabled unsupervised or supervised classification of the scene. The supervised classification enables automatic scene interpretation as for example, discover new urban areas, estimation of the forest height, soil moisture estimation, detection of moving targets, target recognition, structure recognition, 3D mapping, surface-sliding monitoring, etc.
Most important scientific results
Final report,
complete report on dLib.si
Most important socioeconomically and culturally relevant results
Final report,
complete report on dLib.si
