Multiresolution statistical techniques for synthetic aperture radar speckle reduction and image registration.

Abstract

With the advent of the tremendous development in multifrequency, polarimetric and interferometric techniques in synthetic aperture radar (SAR) imaging systems, the need for developing efficient and robust computerized automatic information extraction algorithms is becoming increasingly important for the utilization of SAR imagery. Due to the multi-resolution nature and time-frequency localization property of wavelet representations, image processing in the wavelet domain becomes more efficient and effective. In this dissertation, we develop multi-resolution algorithms for SAR image processing problems, speckle reduction and image registration in particular. Speckle is generally modeled as multiplicative random noise. New techniques based on wavelet multi-resolution analysis have been recently applied in the area of SAR image processing. In some specific cases, these techniques call for a signal model with additive white Gaussian noise and a logarithmic transform is traditionally used to adapt the SAR signal to this requirement. In the first part of this dissertation, the statistical properties of the log-transformed speckle noise are studied. The characterization of the transformed speckle given in this part is meant to call attention to the potential problems of such an approach, and to help in formulating the proper and quantitative justification for this passage. Speckle reduction is a vital first step for most SAR image processing tasks. In the second part of this dissertation, two wavelet despeckling algorithms are developed. The first one fuses the wavelet Bayesian estimation technique and Markov random field modeling to achieve spatially adaptive filtering. In contrast with the first method, the second is designed to work directly on the original SAR image without performing a logarithmic transformation. It accomplishes speckle reduction via the minimum mean square error estimation technique in the wavelet domain. Experiments indicate that both of the two proposed algorithms provide considerable improvement over the other relevant algorithms. In the last part of this dissertation, the potential of mutual information as a similarity criterion for automated SAR image registration is investigated. In order for mutual information to be applicable for SAR images, a couple of remedies are proposed, consisting of speckle reduction prior to registration and a cautious selection of the number of bins for histogram construction. In addition, a multiresolution framework is devised for SAR image elastic registration. An example of registering a pair of real SAR images demonstrates the use of mutual information as a registration quality measure along with the proposed remedies.