Research Of Statistical And The Hierarchical Model In SAR Image DesjDeckling

In the process of image transmission and transformation, it is polluted easily by thenoise which could be produced by the system itself or the external environment. Inorder to interpret and process the real image, it is necessary to remove the noise. Thedistribution of the noise varies with the different kinds of images which have differentimaging principles. Therefor, the nosie suppression can be processed by gettingdifferent noise statistical distribution. The noise of natural image meets a normaldistribution with zero mean. But the distribution of the noise of SAR image is morecomplex. This work was supported by the National Natural Science Foundation ofChina (No.60971128) and Huawei innovation research project (No. IRP-2011-03-04),starting with analysing the distribution of SAR image, combining with the fullhierarchical model to reduce the noise in SAR image. The main work of this paper canbe summarized as follows:(1) After analysing and constrasting the performance of the K-SVD algorithm andthe non-parametric Bayesian dictionary learning algorithm of the the noise suppersionof natural images, it is found that the nonparametric Bayesian dictionary learningalgorithm can be done well for the natural image without restricting the sparsity and thevariance of noise, but for SAR image, it is undesirable. This is because of that the noiseof SAR image doesn’t meet a normal distribution with zero mean. Using local statisticdistribution characteristic of the noise of SAR image, this paper improves the fullhierarchical model of the non-parametric Bayesian dictionary learning algorithm.Supposed that the likelihood function is the multivariate Gaussian distribution, thepaper gets the update formula of parameters in SAR hierarchical model with Gibbssampling and conjugate distribution. The denoised image can be obtained based thisupdate formula. As the amplitude SAR image can’t meet the demand of the conjugatedistribution, this method can’t be suitable for the the amplitude SAR image.(2) As (1) can only be used for intensity SAR image because of the restriction ofconjugate distribution,(2) use maximum posterior probability to update the SARhierarchical model. The experiment of this method shows this method can reducespeckle in extended uniform regions and keep edge well, furthermore it prevents theGibbs-like ringing happening. Compared with (1), it is more pervasive.(3) In order to keep the edge and point target information, this paper use a pixel classification algorithm before noise-suppression algorithm. The pixel classificationalgorithm for SAR image divides one SAR image to three SAR images: edge image,texture image and homogeneous image. Applied (1) and (2) to reduce noise this threeimages, the dictionaries of this three images various, the edge image gets the smallestone and the homogeneous image gets the biggest one. After despeckling, a clean SARimage will be produced by adding the three result images together.