Despeckling Of SAR Imagery Based On Markov Random Field

Synthetic Aperture Radar(SAR) is a kind of high-resolution imaging radar which has abilities of all-weather, multi-polarization, multi-frequency, multi-angle of depression and penetration through smog, cloud, rain and vegetation. SAR has developed very fast in recent years and is increasingly widely used in many fields such as battlefield scouting, aero-photographing, terrestrial remote sensing, terrestrial resource exploration, mapping, and disaster forecast, etc.Despeckling is one of the most basal and important studying field in SAR image processing. The difficulty lies on how to preserve structures while remove speckles sufficiently. Despeckling methods such as statistical filters, multi-scale approaches and Bayesian approaches all can be regarded as making tradeoff between despeckling and structure preserving. There doesn't exit an approach which can do perfectly in both aspects.Markov Random Field(MRF) is successfully used for modeling spatial correlations of pixels. Despeckling based on MRF is a hotspot in the field of speckle suppression. In this thesis, new ideas and approaches of Bayesian despeckling both in image and wavelet domains based on MRF are presented. The emphasis of the work lies on solving the aforementioned conflict, that's to preserve structures while remove speckles.An MRF model called Structure-Preserving MRF is proposed. SPMRF can adjust its weighting parameters adaptively according to image's local characteristics, which overcomes ago MRFs' invalidation on modeling structures. And despeckling based on SPMRF achieves satisfying results. A Bayesian approach based on MMRF model with adaptive neighborhood is also proposed which takes advantages of MMRF's low computing loads, high speed and sufficient speckle removal and puts away MMRF's over-smoothing of structures with fix-neighborhood.In wavelet domain, estimate of wavelets' hidden states combining Hidden Markov Tree and Hidden Markov Random Field is proposed, in which inter-scale and infra-scale correlation of the coefficients are utilized efficiently. This leads to more accurate estimation of hidden states. With states known the Bayesian estimation is applied to the coefficients to eliminate noises' infection. Experiments show that the despeckling method based on HMT-HMRF model can not only despeckle sufficiently but also preserve structures well.