SAR Image Despeckling Algorithm Via Sparse Representation And Nonlocal Means

Synthetic Aperture Radar(SAR) is a kind of active remote sensing observation system which is widely used in both military and civil field. The unique properties of electromagnetic make SAR imaging approach without being affected by the influence of light and weather, thus SAR imaging system can work with 24 hours, all-weather condition. With the development of computer technology, the automatic targrt recongization method has been used in more and more occations. Hovever, for conherent imaging mechanism, the observed SAR image is contaminated with a serious speckle noise which severly damage the useful image content and bring much difficult to the reconization and interpretation of SAR image. Therefore, it is very necessary to conduct SAR image despeckling.Although various kinds of despeckling algorithms have gained great development in recently more than 30 years, it is difficult to achieve balance between the suppress of speckle and the maintain of edge details, structures, and radiation properties of SAR images using these despeckling algorithm. Therefore it is necessary to design more powerful algorithm to despeckle SAR images. Given the great progress achieved by the theory research of sparse representation and the successful application of nonlocal means in image processing, we propose a despeckle algorithm and its improved version based on the sparse representation and nonlocal means method under the basis of analyzing speckle statistical property. The author’s major contributions are summerized as follows:Analyzing the sparse representation and nonlocal means method which are widely used in image processing application in recent years, and considering the excellent performance of Nonlocal Centralized Sparse Representation(NCSR) model applied in nature image processing, we extend this model to SAR image despeckling. Given the big difference between the multiplicative speckle noise of SAR image and the additive Gaussian noise of nature image in terms of statistical distribution of noise characteristics, we firstly make some preliminary experiments and analyses to test whether the SAR image is suitable for the model or not. The logarithmic SAR image is divided intooverlapping sub-image patches which are used to learn an overdictionary using K-means clustering algorithm and principle component analysis algorithm, and then adaptive sparse coding are applied in these patches via using the learned overdictionary. At the same time, the estimation of sparse coding coefficients corresponding to clean patches is computed by nonlocal means method, and the sparse coding coefficients of denoising patches is computed by iterative threshold shrinkage algorithm. After returning to logarithmic domain via inverse transformation, the bias is corrected in logarithmic SAR despeckling image to improve the despeckling ability.Considering that the noise of logarithmic SAR image is not strictly meet Gaussian distribution, it is not appropriate to apply the Gaussian similarity measure used in NCSR model to SAR image processing in logarithmic domain. After discussing the noise statistic distribution properties of logarithmic SAR image, we derive a new similarity measure based on the generalized likelihood ratio criterion to replace the Gaussian similarity measure. Experiment prove that this improvement achieve better performance, especially in low look numbers of SAR images.For the proposed SAR despeckling algorithm in this thesis, we use several evaluation criteria and comparing algorithms to test the ability of despeckling and the retention capacity of the details information of SAR images. Experiment results demonstrate that our proposed algorithms are effective in the test of the simulated speckle images and the real SAR images.