基于混合模型和水平集的高分辨SAR图像分类

发布时间：2019-05-11 01:22

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)可以全天时和全天候地对地球表面地物成像,而且可以穿透地球的表面。SAR可以获得较高的空间分辨率,因此高分辨SAR图像在军事、农业和医学领域起着越来越重要的作用。但是,高分辨率的SAR图像由于存在比较严重的乘性相干斑噪声,用传统的分类方法对其进行分类,将得不到较好的分类结果。因此,本文提出了能正确描述SAR图像统计特性的基于K-SVD训练字典的混合模型,以及基于混合模型和改进的水平集的SAR图像分类方法,具体的改进思路如下:(1)SAR图像由于存在严重的乘性相干斑噪声,传统的的单模型已不能对高分辨率的SAR图像进行较精确地统计建模。本文提出了一种基于K-SVD算法训练字典的混合模型统计建模方法,此混合模型是基于对数正态分布和威布尔分布两个模型建立的。由于传统的EM算法对SAR图像混合模型的建模过程比较复杂,提出了通过K-SVD算法来训练字典。在本文中,选择基于梅林变换的SKS参数估计方法对对数正态模型、威布尔模型进行参数估计。通过对SAR图像的匀质区域、不匀质区域和极不匀质区域进行混合模型的拟合表明,混合模型能够对各种地物都能进行较好的统计建模。(2)由于基于Gamma统计模型的水平集分类方法不能对高分辨SAR图像进行较高精度的分类,提出了一种基于K-SVD训练字典的混合模型和改进的CV模型水平集的高分辨SAR图像分类方法。在水平集的典型模型Chen-Vase(CV)模型中,假设SAR图像的每个地物区域都具有一致的的强度,实际上SAR图像中存在许多不匀质区域,因此将CV模型用于SAR图像分类存在了一定的局限性。因此,本文提出了在水平集分类方法中,用基于K-SVD训练字典的混合模型来对高分辨率的SAR图像的不同区域进行统计建模。(3)由于边缘信息是SAR图像分类的重要依据,提出了一种将边缘信息和改进的区域信息相结合的高分辨SAR图像水平集分类方法,在区域能量函数中用能较好地对SAR图像进行统计建模的混合模型来代替CV模型中的高斯模型。实验结果表明,本文提出的基于K-SVD算法训练字典的混合模型和改进的水平集分类方法可以对高分辨率SAR图像进行较高精度的分类。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture Radar,SAR) can image the earth's surface objects all day and 24 hours a day, and can penetrate the earth's surface. Sar can obtain high spatial resolution, so high resolution SAR images are military. Agriculture and medicine are playing an increasingly important role. However, because of the serious multiplicative coherent speckle noise in high resolution SAR images, the traditional classification method will not get good classification results. Therefore, this paper proposes a hybrid model based on K-SVD training dictionary, which can correctly describe the statistical characteristics of SAR images, and a SAR image classification method based on hybrid model and improved level set. The specific improvement ideas are as follows: (1) because of the serious multiplicative coherent speckle noise in SAR images, the traditional single model can no longer accurately model the high resolution SAR images. In this paper, a hybrid model statistical modeling method based on K-SVD algorithm training dictionary is proposed, which is based on two models: lognormal distribution and Weibull distribution. Because the traditional EM algorithm is complex in the modeling process of SAR image hybrid model, a K-SVD algorithm is proposed to train the dictionary. In this paper, the SKS parameter estimation method based on Merlin transform is selected to estimate the parameters of lognormal model and Weibull model. The mixed model fitting of homogeneous region, uneven region and extremely uneven region of SAR image shows that, The hybrid model can carry on the better statistical modeling to all kinds of ground objects. (2) because the level set classification method based on Gamma statistical model can not carry on the high precision classification to the high resolution SAR image, A high resolution SAR image classification method based on K-SVD training dictionary hybrid model and improved SAR model level set is proposed. In the typical model Chen-Vase (CV) model of the level set, it is assumed that each feature region of the SAR image has the same intensity. In fact, there are many uneven regions in the SAR image. Therefore, there are some limitations in the application of CV model to SAR image classification. Therefore, in the level set classification method, a hybrid model based on K-SVD training dictionary is proposed to statistically model the different regions of high resolution SAR images. (3) because edge information is an important basis for SAR image classification, In this paper, a high resolution SAR image level set classification method is proposed, which combines edge information with improved region information. In the regional energy function, a hybrid model which can well model SAR images is used to replace the Gao Si model in CV model. The experimental results show that the hybrid model based on K-SVD algorithm training dictionary and the improved level set classification method can classify high resolution SAR images with high accuracy.
【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】