The Study Of SAR Image Despeckling Algorithm Based On Directionlet Transform And Fast Implementation

Since synthetic aperture radar (SAR) can obtain the information of objects over and under earth surface under any weather condition and any illumination condition, SAR images are widely used in many fields, such as ocean monitoring, resource and geological exploration, observation and evaluation of corps growing, military affairs and etc. However, due to the application of the interference principle in the imaging systems, speckle inherently appears in the obtained SAR images. Speckle not only severely degrades the quality of SAR images and conceals the fine structures, but also makes difficulties in post-processing tasks, such as image segmentation, object recognition, object classification and etc. Thus, it is very important to remove the speckle noise of SAR images.The traditional two-dimension wavelet base functions lack multidirectionality and anisotropy, this leads to a nonsparse representation for images. The anisotropic one-dimension features of images play an important role in the human visual perception, however, a number of wavelet coefficients intersects at the discontinuous edges when we use the two-dimension wavelet to decompose the images, and leads to a nonsparse representation. As a multiscale geometrical analysis tool, directionlet transform is more sparse than wavelet transform, and has many excellent properties such as multiple scale, multiple direction and anisotropy. Thus, directionlets can capture the anisotropic one-dimension features of image and achieve a sparse representation for images.Since different surface features lead to speckle noise with different statistical characterizations, it is very difficult to establish model for speckle using a fixed statistical model. Based on the study of directionlet transform, we first discussed the statistical properties of directionlet coefficients of SAR images and employed the Cauchy distribution, Gaussian mixture distribution and Laplacian mixture distribution to fit the directionlet coefficients. Then, a maximum a posteriori estimator is derived within the framework of Bayesian theory. Finally, the speckle noise was removed by using the estimator. In order to derived the corresponding maximum a posteriori estimator, we estimated the parameters involved in the mixture distribution using the expectation maximization algorithm. As a sparse representation tool, the directionlets remove most of the relations existing between the pixels, but relatively significant dependencies still exist between directionlet coefficients, which affect the performance of the image processing algorithms. After studying the dependency between directionlet coefficients, we designed a linear predictor to introduce the dependency into the maximum a posteriori estimator in terms of mutual information. The linear predictor enhanced the despeckling performance of the proposed algorithms. The experimental results show that the proposed method based on directionlet transform preserves the edge features while removing speckle noise of SAR images.Aiming at the characteristics of SAR image with massive data and high computational complexity of the despeckling algorithms, we studied the hardware structure of graphics processing unit (GPU) and compute unified device architecture and realized the parallel algorithm design of SAR image despeckling algorithm based on thresholding. To achieve this goal, we placed the portion with high parallelism and intense computation on GPU and put the portion with strong logicality on the Central Processing Unit (CPU). The experimental results show that the executing efficiency of the algorithm based on GPU is obviously better than the one based on CPU. If the computing time without containing copy operation, then the executing efficiency of the parallel algorithm will be much better.