SAR Image Labeled Algorithm Based On Markov Random Fields

Synthetic aperture radar(SAR)has potential advantages in image label field(compared with the optical sensor),as they are independent of optical condition and atmosphere.More and more researchers pay attention to the SAR image-processing algorithm,especially in SAR image segmentation,change detection and classification.Recently,the methods based on Markov theory have been widely used in SAR image-labeled algorithm.Most of these methods use the spatial contextual information within Markov random fields(MRFs)to enhance the robust against noise.It has been proved that these kinds of method can reduce the speckle noise of SAR image obviously,especially for the homogenous data.Besides,MRFs are unsupervised algorithm,so it is an effective method when there is no label truth.Thus,MRFs takes advantages of both anti noise and free from label truth.This dissertation deals with SAR image label problem.We analyze the main issue within the MRFs in image processing,and propose new methods for SAR image segmentation and change detection.In SAR image segmentation,we analyze the reason why the edges are absence,and construct a correlated evidential Markov model that treats mixture Gaussian model as it possibility part.For SAR image change detection,we discuss how to use MRFs to reduce noise and keep the edges meanwhile.The main content is as follows:1.Considering both interpretations simultaneously,we propose a Markov field model called Gaussian mixture-hidden evidential Markov field(GM-HEMF)which integrates,on the one hand,an auxiliary field U defined in the evidential domain related to the prior field X to model its unreliability,and on the other hand,a second auxiliary field U~?,modeling the unknown form of noise distributions p(ys|xs).For parameters estimation,we use iterative conditional estimation,whereas maximization is performed through iterative conditional mode.The performance of the proposed model is assessed against the original EMF on real SAR images.2.To generate reasonable difference data for the change detection,we propose novel feature-level difference data,namely,multicontextual MI data(MMID),based on the bivariate Gaussian distribution(BGD)for SAR image change detection.The proposed MMID are constructed based on the quadrilateral Markov random field(QMRF)and can be factored into the linear combination of the entropies.Thus,MMID are able to capture the intertemporal dependence of the observed data.Then,we construct the 2-D entropies based on the BGD.It is well acknowledged that the Gamma distribution functions well in modeling the SAR data.However,the probability density function(pdf)of Gamma is more complicated than the Gaussian distribution.The bivariate Gamma distribution is much more complicated than the BGD.Thus,in our dissertation,we adopt the BGD to fit the two original images.By utilizing the BGD to model t he 2-D pdf in the computation of the entropy within the two temporal image patches,MMID are suitable for the change detection.Then,due to the capture of the statistical properties of the two temporal images,MMID can be taken as the feature-level difference data rather than the pixel-level data.3.We propose a novel semisupervised SAR images change detection algorithm using discriminative random fields based on maximum entropy principle(MEDRF).In the context of MEDRF,two generative models,named as bias model and correction model,are both constructed based on MRF model to capture the spatial-contextual information,but trained by labeled samples and unlabeled samples,respectively,to take advantage of the labeled information in addition to the unlabeled information.Then,we derive two constraints from the two generative models.Thus,through maximizing the entropy of target discriminative distribution subjected to the derived constraints,the achieved MEDRF is a fusion of the bias model and the correction model.The MRF-based correction model can mitigate the impact of bias on the detection performance associated with the MRF-based bias model.In this way,MEDRF model is able to capture the spatial-contextual information of labeled samples and unlabeled samples,and benefits from the generative approach and the discriminative approach in the change detection.4.We deal with the change detection from simultaneous segmentation of current and past images using Markov fields and Dempster-Shafer theory of evidence(TE).The novelty is to apply new?double evidential Markov fields to segment the couple(current image,past image),which gives the change detection.Parameter estimation method is also specified,resulting in unsupervised change detection.5.We propose a mean-ordered and quadrilateral MRF(MO-QMRF)based method for SAR image change detection.Considering that spatial information is of great importance in enhancing the robustness against noise,and SAR images are badly interfered by speckle noise,we adopt QMRF,which includes eight-neighborhood information but can be decomposed to the product of 2 dimensional pdf,to construct the joint pdf between the bitemporal images.The parameters are estimated in a traditional way to keep the resolution of the original data.Thus the proposed method can enhance the robustness against noise as well as detect the small punctual changes.Given the false alarm rate(FAR)detection threshold,the only left parameter,threshold can be selected automatically.MO-QMRF compares two temporal means,and thus it can be extended to time series change detection easily.