SAR Image Despeckling Baesd On Sparse Representation

SAR images have been widely used in plenty of fields of national security and economy. Due to the inherent drawbacks of imaging mechanism, the speckle phenomenon is always companying with SAR images. The presence of speckle make the interpretation and understanding of SAR images become difficult. Therefore, speckle reduction is a key and essential technology in SAR image processing. In recent years, there has been an increasing interest in the search for sparse representation of signals. A signal can be represented as a sparse linear combination of some atoms under a specific dictionary. Sparsity has been successfully used for image compression, inpainting and more. Based on the widely concerned sparse representation, the paper focuses on the suppression of speckle noise in SAR images. The major works of this paper consist of the following three parts:(1) A method is proposed for reducing the SAR image speckle noise based on sparse representation. Firstly, when representing an image patch, we introduce the size of the patch and the pixel information as the control factor to decrease the effect of the speckle noise. Secondly, using the KSVD algorithm, an adaptive dictionary is trained which represents the image content effectively from the noisy SAR image patches. For each pixel, the estimated values are then averaged to obtain the processed pixel. At last, the nonlinear anisotropic diffusion is employed to deal with the difference image to enhance the point targets in the reconstructed image. Experimental results demonstrate that the proposed method has a good despeckling performance while preserve the edge clearly.(2) An approach is presented based on sparse dictionary (S-KSVD) model for speckle reduction in SAR images. Because the dictionary trained by the KSVD algorithm is highly structured, with noticeably regular atoms, the S-KSVD model is introduced. Applying the signal-dependent additive noise model for SAR image and combing the different format, we derive the sparse coding stage for SAR image in intensity and magnitude format. Then using the S-KSVD model, a sparse matrix and the updated representation coefficients are obtained, which are used to reconstruct the despeckled image. The method can effectively reduce the speckle noise in homogeneous regions, with the radiometric preservation almost perfect.(3) An algorithm is addressed according to multiscale dictionary learning in wavelet domain for SAR image despeckling. The signal-dependent additive noise model is extended to image patch in stationary wavelet domain. Assumed each image patch been a homogenous region, the additive noise is modeled with the Gaussian distribution. After that, we train the multiscale dictionary based on KSVD algorithm in stationary wavelet domain. Then each wavelet subband is reconstructed by its sub-dictionary and representation coefficients. To preserving edge information, we segment the edge region in original SAR image and obtain the edge index matrix using the variance image. The original wavelet subband and reconstructed subband are fused together by utilizing the edge index matrix to generate the modified wavelet subband. The method combines the explicit dictionary with implicit dictionary naturally, which make the despeckling performance superior to many classical methods, with information details well preserved.This work was supported in part by the National Natural Science Foundation of China under Grant no.60971128,61072106,61173090and the Fundamental Research Funds for the Central Universities under Grant no. JY10000902001.