Study On Despecking And Segmentation Methods Of SAR Image

Synthetic Aperture Radar (SAR) can obtain high range-resolution by transmittingwideband radar signal and high azimuth-resolution by long synthetic aperture.Therefore, high-resolution SAR image with large-size can be captured. SAR has theability to image the earth’s surface in nearly all weather conditions for long distance.Together with its high spatial resolution, the SAR has a more and more important role atfields of geosciences, hydrology and bionomics, etc. While, the subsequent SAR imageunderstanding is difficult because of the presentation of SAR speckle. Therefore, interm of the observation value of the SAR image, the estimation of the true value of theRadar cross section (RCS) is an important research topic of SAR image pre-processing.These years, the subsequent SAR image segmentation and change detection have moreand more attentions in the Radar signal processing fields.In this dissertation, the SAR image despeckling technology, SAR imagesegmentation and change detection methods are studied, and a series of practicaleffective methods are proposed. The author’s major contributions are outlined asfollows:1. The non-local despeckling method for SAR image based on anisotropicGaussian directional window and Stein unbiased risk estimation (SURE) aggregationhas been proposed.Aimed at the shortage of similar region and directional information capture lack forSAR image despeckling using conventional non-local means method (NLM), a newNLM SAR image despeckling method is proposed based on anisotropic Gaussiandirectional window and Stein unbiased risk estimation (SURE) aggregation. Thismethod is based on multiple different directional anisotropic Gaussian windows, whichcan match the local geometric structure and can capture more pattern redundancy thanthe square window in the conventional NLM. Then, the ratio measurement strategy isutilized to compute the similarity of two patches. Finally, the results of NLM withdifferent anisotropic Gaussian windows are aggregated using the Stein unbiased riskestimation criterion. For multiple SAR images, the experiment results show that the newmethod has advantages in the SAR image despeckling performance, and can wellpreserve the local geometric structure information, which is essential for understandingand interpretation of SAR image.2. Spatial threshold-segmentation method of SAR image based on adaptiveGaussian weighted directional window and three-dimensional Otsu method has beenproposed.Aimed at the shortage of the abilities of noise removing and small targetpreservation for the conventional three-dimensional (3D) Otsu thresholding method, a new three-dimensional Otsu method based on adaptive Gaussian weighted directionalwindow is proposed. The new method improves the window setting method of the3DOtsu. The window size, scale and filtering angles are adaptively determined by the localstationarity characters. Then, based on the non-local multiple directions similaritymeasurement, the pattern redundancy in the image can be captured effectively. Finally,the3D histogram is constructed based on the gray value, weighted mean value andweighted median value. And the threshold vector is computed by the maximumbetween-class variance method to segment the image. The proposed method has thebetter segmentation performance, with better performance for noise removal and smalltarget preservation.3. SAR image segmentation method in overcomplete Brushlet domain based onGray-Level Cooccurrence Probability and Fuzzy C-Mean (FCM) clustering has beenproposed.Aimed at the shortage of edge preservation and low direction-resolution for SARimage segmentation based on the conventional wavelet transform domain, a newsegmentation method is proposed based on Gray-Level Cooccurrence Probability(GLCP)features in the overcomplete Brushlet domain. This method compresses the redundantGLCP features extracted by the adaptive window Gaussian filtering in differentdirection coefficient blocks using compressed sensing, then the Fuzzy C-Mean (FCM)clustering method is utilized to complete the clustering and obtain the segmentationresult. The experiment results show that the new method has advantages in the edgepreservation and direction extraction, and obtains better segmentation results withrespect to other methods.4. SAR image segmentation method based on overcomplete Brushlet transform andspectral cluttering ensemble has been proposed.Aimed at the shortages of edge preservation and low direction-resolution based onthe conventional wavelet transform domain and the sensitivity of the scaling parameterof the spectral cluttering, a new segmentation method is proposed based onovercomplete Brushlet transform and spectral cluttering ensemble. This method solvesthe sensitivity problem of the spectral cluttering and low direction-resolution of thewavelet transform. This method has better performance than the conventional spectralcluttering, and avoids the selection problem of the scaling parameter. The new methodhas better segmentation performance than the conventional methods, and hasadvantages in the preservations of the edge and direction details.5. SAR image change detection method in overcomplete Brushlet domain based onStein unbiased risk estimation has been studied.Aimed at the shortage of similarity character capture and low direction-resolutionfor SAR image change detection and the difficult for the modeling of speckle noise, a new2D-otsu SAR image change detection method is proposed based on Stein unbiasedrisk estimation (SURE) and linear expansion of threshold (LET) in the overcompleteBrushlet domain. Based on the diversity image, this method combines the localanisotropic Gaussian weighted nonlinear mean procedure in the overcomplete Brushletdomain and linear combination with the minimum mean squared error in the originaldomain to obtain mean character after the speckle noise is removed. Then, changedetection is processed with combining the mean character and gray-level character. Theovercomplete Brushlet resolves the problem of low direction-resolution, and canaccurate positioning the texture of each direction, frequency and position. The SUREand LET principles together give a fast and efficient algorithm that only solves a linearsystem of equations, without the prior knowledge of the speckle noise model. The newmethod has advantages in the change detection performance, and can well preserve thedetailed information such as the texture edge.