SAR Image Despeckling Based On Sparse Representation And Low-rank Approximation

Synthetic Aperture Radar(SAR) system has the characteristics of all-weather, all-time, high resolution and strong power in penetrating, so SAR has been widely used in both military and civilian aspects. However, because of its microwave coherent imaging, SAR image is corrupted by speckle noises which greatly reduce the resolution of SAR image and increase some difficulties for the following processing and interpretation. So how to suppress speckle noise existed in the SAR image is very important. Analyzing the model and statistical properties of the speckle noise and combining sparse representation and low rank approximation, several novel SAR image despeckled schemes are proposed in this paper. The main task of this paper can be summarized as the following three aspects:1. A novel algorithm based on clustering and Improved Dictionary Lear ning is proposed, which is used for speckle noise removing. According to the fact that there are many image patches having similar structures in an image, K-means clustering algorithm is performed to achieve them grouping. In order to tap much texture information inherent in SAR image, we use principal component analysis for image patches which have the same cluster to build the corresponding PCA dictionary containing rich texture information. At last, Initialize dictionaries with PCA dictionaries, use Improved Dictionary Learning to get the resulted image.2. A novel SAR image despeckled scheme via clustering-based sparse representation with SSIM proofreading is proposed. Using the multi-directional and anisotropy advantages of Directionlet transform, a speckle noise estimation algorithm based on Directionlet transform is used to estimate speckle noise existing in SAR image. According to large number of patches having similar structure inside the image, both within the same scale, as well as across different scales, Directionlet transform is used to obtain similar pathches of multi-scales images and then the K-means algorithm based on SSIM proofreading is performed to achieve similar patches grouping in this paper. Finally, use the Clustering-based Sparse Representation to sparse representation and then achieve the speckle reducted image.3. An improved algorithm based on spatially adaptive iterative singular-value thresholding algorithm is proposed in this paper. Though the analysis of low rank characteristic of similar patches, Singular Value Decomposition is used to low rank approximation of SAR image to remove the speckle noise in SAR image. In order to retain much texture information of SAR image, enforce the gradient distribution of the denoised image to be close to the estimated gradient distribution of the clear SAR image in the process of SAR image denosing. The experimental results demonstrate that the proposed algorithm can achieve convincing improvement over previous state-of-the-art methods in terms of both target and edge preservation.