Research On SAR Image Despeckling Based On Multiscale Geometric Analysis Technology

Synthetic aperture radar (SAR) images are inherently affected by coherent speckle noise which significantly degrades the image quality and increases great difficulties for SAR image interpretation. Removing noise from SAR image called despeckling is a key pre-processing step for post processing for SAR image. The multi-resolution analysis performed by the wavelet transform has been proved to be a powerful tool. Because of running short of multi-directional character, wavelet transform is not good at describing images'edge and contour, a lot of transform tools with multiscale and multidirectional characteristics are emerged as the times require, such as Curvelet transform, Brushlet transform, Contourlet transform and Nonsubsampled Contourlet transform etc. This paper combining the multiscale and multidirectional characteristics of Nonsubsampled Contourlet transform and the specialty of SAR image speckle noisy, studies SAR image despeckling methods. The innovative points of this paper are as follows:1. An improved Nonsubsampled Contourlet Transform (NSCT)-based method has been proposed, using subbands mask prior models and directional information for synthetic aperture radar (SAR) image despeckling. We use modified logarithm gaussian distribution to approximate the histogram of coefficients representing important information, and use mixed exponential distribution to fit the histogram of coefficients representing unimportant information. Depending on Bayes principle, we obtained shrinkage factor to modify nonsubsampled Contourlet decomposed coefficients of SAR image. The results show that our shrinkage algorithm achieves great outcome for SAR image despeckling.2. The second innovation of this paper is to add in the operation of mathematical morphology for despeckling. This method uses mathematical morphological open operation to improve binary masks of subband coefficients and designs special structural element according to multiscale and multidirectional characteristics of nunsubsampled Contourlet transform. The experiment results improve that the decpeckling method combining mathematical morphological removes small abrupt caused by speckle noise, therefore it gets approving despeckled effects.3. It is an improved Wiener filtering algorithm which is special because of double Wiener filtering and a couple of thresholds for image denoising. We use Generalized Gaussian Distribution (GGD) to fit the histogram of subband coefficients and use the result of the first step of Wiener filtering to estimate coefficient variance which is a significant parameter for the second step of Wiener filtering. The experiment results show that this method improves peak signal to noise ratio for denoised nature image and promotes equivalent number of looks (ENL) for despeckled SAR images.