The Study Of Parallel Algorithm For SAR Image Despeckling In Directionlet Domain Based On GPU

Synthetic Aperture Radar(SAR)is a kind of new radar technology,which has been widely used in many fields.It is all-weather,that is,SAR imaging is not affected by bad weather such as wind,rain,thunder and fog.However,there is speckle noise in SAR image because of its imaging principle,which seriously affects the quality of SAR image and also brings trouble to subsequent image processing and analysis work.With the continuous development of SAR imaging technology,we can get many high resolution SAR images.Moreover,in the application of SAR imaging technology,most of them need to be continuously imaged,so the amount of data obtained is extremely huge.Therefore,efficient denoising algorithm is the key to SAR image research nowadays.In the existing SAR image denoising algorithms,almost all are traditional serial algorithms.With the rapid development of GPU technology,GPU computing power is far beyond serial computing.Thus,it is of great practical significance to study the parallel algorithm for speckle noise suppression in SAR images.In this dissertation,a parallel SAR image denoising algorithm is implemented on GPU by combining the directionlet transform,one kind of the mature third generation wavelet transform.The reason why the standard two-dimensional wavelet transform is not used here is that it uses the standard orthogonal basis,which makes it isotropic.It can not capture and express the information of the image in the edge and contour well,which will seriously affect the follow-up work of SAR image.Thus,the anisotropic directionlet transform is selected in the paper,which not only makes the contour and edge of image well expressed,but also makes the transformation directional.In other words,directionlet transformation is no longer confined to horizontal and vertical directions,but can be transformed in any given direction.According to the sparse characteristic of directionlet transform coefficients,they are modeled and solved in this paper.Then,the noise-free SAR image is estimated from the noisy SAR image using the parameters obtained above,which achieves the effect of noise removal.There are many parallelizable processes in the processing of noise removal and each process is parallelized for computation based on GPU in this paper,which complete the parallelization of the whole algorithm and improve the efficiency of SAR image denoising.And the following three aspects are mainly included in this SAR image denoising.(1)GPU-based directionlet transformation.In this part,the generator matrix is first determined,and then the SAR image is decomposed into different sub-images according to the generator matrix.In this process,computer memory is allocated for the input SAR image and the cosets to be output respectively.According to the corresponding relationship between the input and output,we establish the kernel function and realize the parallelization of the decomposition algorithm by calling it on the GPU.Then anisotropic wa'velet transform is applied to each coset image.The sub-function of the most important convolution process is established which is called and calculated on GPU.The outer loop and kernel function are also established and AWT is realized by calling convolution sub-processes.In this way,the directionlet positive transformation is completed on GPU,which can get the high and low frequency information of the original image and greatly improve the efficiency of the transformation.(2)Parameter estimation based on GPU.According to the statistical characteristics of directionlet transform coefficients,we model the high frequency coefficients after transformation,and use GMM to fit their statistical curves.In order to solve GMM parameters,we adopt the classical expectation maximization(EM)algorithm.In this part,we only use GPU to calculate the parameters because that EM algorithm itself can not be parallelized.The large sub-image is decomposed into small image tiles and their parameters are solved separately.Then take the average value of these parameters which can improve the accuracy of the model parameters,so that the parameters solved can better fit the original coefficient statistical curve.As for the statistical model of noise,we assume that the noise is a Gaussian noise with an unknown variance and a mean value of zero.For noise variance,we use standard deviation estimation method.First,we build a sort subroutine which is calculated and invoked on GPU.After the data is loaded,we group the data and call the sort subroutine to find the median of each group.Then we reload the median into a new data and cycle to get the result.(3)SAR image reconstruction based on GPU.According to the obtained model parameters and noise parameters,we estimate the noise-free SAR image from the noisy image by the method of maximum a posteriori probability estimation,so as to achieve the purpose of denoising.In this process,we load the SAR image and parameters with noise.The kernel function is the estimation method.Each thread contains a pixel,and the noise-free image is obtained by calling the kernel function.These estimated images also need directionlet inverse transformation to reconstruct noiseless original images.The final experimental results show that the proposed GPU-based parallel SAR image speckle noise suppression algorithm in directionlet domain can not only maintain the accuracy of the original serial algorithm,but also be very efficient and the running time of the program can be greatly reduced.