基于线性最小均方误差估计的SAR图像降噪

针对合成孔径雷达(synthetic aperture radar, SAR)图像降噪过程中容易引起细节纹理信息损失的问题,该文结合SAR图像相干斑噪声的统计特性,提出了一种基于变换域系数线性最小均方误差(linear minimum mean-square error, LMMSE)估计的SAR图像降噪方法。首先通过SAR场景下的Kmeans聚类算法将相似图像块聚类;然后针对每一类相似图像块集合进行奇异值分解(singular value decomposition, SVD),得到同时包含图像块集合行列相关信息的含噪奇异值系数;为从含噪奇异值系数中更准确地估计出真实图像奇异值的系数,先通过加性独立信号噪声(additive signal-dependent noise, ASDN)模型将乘性噪声转化为加性噪声,再利用LMMSE准则对奇异值系数进行估计,最后将估计结果重构得到降噪后的图像块集合。实验结果表明,该方法充分利用相似图像块集合奇异值系数稀疏的特性,采用LMMSE准则估计奇异值系数,既保证了系数中噪声分量的去除又避免了图像纹理细节对应小系数的丢失,不仅去噪效果明显,同时能有效地保持图像纹理细节,具有良好的图像视觉效果。

SAR image denoising via linear minimum meansquare error estimation

In order to solve the problem that many detail texture information is lost during the synthetic aperture radar (SAR) image denoising process, SAR image denoising approach based on the estimated transform domain coefficients by the means of linear minimum mean square error (LMMSE) is proposed, which combines the statistical characteristics of the speckle noise in the SAR image. Firstly, cluster image blocks into disjoint sets of similar blocks through Kmeans corresponding to the SAR scene. Secondly, perform singular value decomposition (SVD) for each set of similar blocks, and the noisy singular value coefficients containing the correlation of rows and columns of the set of similar blocks can be obtained. In order to estimate the noise-free singular value coefficients more accurately from the noisy singular value coefficients, the additive signal-dependent noise (ASDN) model is used to convert the multiplicative noise into the additive noise, then estimate the noise-free singular value coefficients by using the LMMSE technique. Finally, obtain the denoised set of similar blocks by reconstructing the estimation results. The experiment results show that the proposed method makes full use of the sparse characteristics of the set of similar blocks, and utilizes the LMMSE technique to estimate the coefficients, which can not only remove the influence of the noise but also avoid the loss of the important texture details of the image and the denoised image has better visual quality.