基于稀疏表示和低秩逼近的SAR图像降斑

发布时间：2018-07-31 17:40

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)具有全天候、全天时、高分辨率和穿透力强等特点,因此,SAR在军事方面和民用方面都得到了广泛的应用。但是,由于SAR系统是微波相干成像,SAR图像在获取的过程中受到了斑点噪声的影响。斑点噪声的存在大大降低了SAR图像的分辨率,影响了后续的处理与解译。因此,如何抑制SAR图像中的斑点噪声非常重要。通过分析SAR图像的斑点噪声模型和统计特性,结合稀疏表示理论和低秩逼近理论,本文提出了几种新的SAR图像降斑算法。本文主要包括以下三个方面:1.提出了一种基于聚类和提升字典学习的SAR图像降斑算法。考虑到图像中具有大量相似结构的图像块,利用K均值聚类算法,构造相似图像块集合。为了充分挖掘图像块中包含的纹理细节信息,本章利用主成分分析法,提取各个相似图像块集合的主成分分量,构造相应的PCA字典。以PCA字典为初始字典,采用提升字典学习算法对相似图像块进行稀疏表示与重构,得到最终的降斑结果。2.提出了一种基于结构相似度校正聚类的稀疏表示SAR图像降斑算法。利用方向波变换的多方向性和各向异性,在方向波域对SAR图像进行斑点噪声估计。考虑到相似图像块不仅仅存在于同一尺度的图像中,不同尺度的图像中也包含大量的相似图像块,本章通过对SAR图像进行方向波变换来获取不同尺度的图像块,并采用一种基于结构相似度校正的聚类算法对图像块进行类别划分。最后,利用基于聚类的稀疏表示算法对每类图像块进行稀疏表示和重构,得到最终的降斑结果。3.提出了基于改进的空间自适应迭代奇异值阈值的SAR图像降斑算法。考虑到相似图像块集合具有低秩性的特点,本章利用奇异值分解对SAR图像进行低秩逼近重构,进而达到降斑的目的。在降斑的过程中,为了更好的保留图像中的纹理信息,以原始图像的梯度直方图作为参考,通过约束更新后图像的梯度直方图来达到纹理增强的目的。该方法不仅能够很好的抑制图像中的斑点噪声,且可以很好的保留图像中的点目标和纹理信息。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture) has been widely used in military and civil fields because it has the characteristics of all-weather, high resolution and strong penetration. However, the SAR system is affected by speckle noise in the acquisition process of microwave coherent imaging. The presence of speckle noise greatly reduces the resolution of SAR images and affects the subsequent processing and interpretation. Therefore, how to suppress speckle noise in SAR images is very important. Based on the analysis of speckle noise model and statistical characteristics of SAR images, combined with sparse representation theory and low rank approximation theory, several new speckle reduction algorithms for SAR images are proposed in this paper. This article mainly includes the following three aspects: 1. A SAR image speckle reduction algorithm based on clustering and lifting dictionary learning is proposed. Considering a large number of similar image blocks in the image, the K-means clustering algorithm is used to construct the set of similar image blocks. In order to fully mine the texture details contained in image blocks, this chapter uses principal component analysis (PCA) to extract the principal components of similar image blocks and construct corresponding PCA dictionaries. Using the PCA dictionary as the initial dictionary, the lifting dictionary learning algorithm is used to sparse represent and reconstruct the similar image blocks, and the final speckle reduction result. 2. A sparse representation SAR image speckle reduction algorithm based on structural similarity correction clustering is proposed. The speckle noise of SAR images is estimated in directional wave domain by using the multi-directivity and anisotropy of directional wave transform. Considering that similar image blocks not only exist in images of the same scale, but also contain a large number of similar image blocks in different scales, this chapter uses directional wave transform to obtain different scales of image blocks. A clustering algorithm based on structural similarity correction is used to classify image blocks. Finally, the sparse representation algorithm based on clustering is used for sparse representation and reconstruction of each image block, and the final speckle reduction result. 3. An improved spatial adaptive iterative singular value threshold algorithm for SAR image speckle reduction is proposed. Considering the low rank characteristic of similar image block set, this chapter uses singular value decomposition to reconstruct SAR image with low rank approximation, so as to achieve the purpose of speckle reduction. In the process of speckle reduction, in order to preserve the texture information of the image better, the gradient histogram of the original image is used as the reference, and the gradient histogram of the updated image is constrained to achieve the purpose of texture enhancement. This method can not only suppress the speckle noise in the image, but also preserve the point target and texture information of the image.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】