基于稀疏表示的面向对象极化SAR图像分类

发布时间：2019-02-12 09:10

【摘要】：极化SAR图像分类作为遥感图像处理与研究的重要方向之一已日益受到人们的关注。由于极化SAR数据中大量存在的噪声点会对图像的进一步处理与研究造成严重影响,因此,如何克服噪声干扰,提取有效特征并进行分类是极化SAR图像分类面临的关键问题。为了解决上述问题,本文引入了面向对象的思想以及稀疏表示理论。面向对象的思想通过初始分割提取对象进行分类抑制了大量噪声点的干扰,而稀疏表示理论对噪声的鲁棒性也提高了分类的正确性,两种算法相结合可以得到很好的分类结果。本文的四个主要工作内容如下所示:1.对极化SAR图像进行了初始分割。本文用于初始分割提取对象的算法将纹理信息与统计区域合并算法相结合,使用了Bhattacharyya距离合并准则与统计区域合并准则,可以获得更为正确的图像分割结果。纹理信息的提取是通过改进后的LBP算子:局部一致性LBP算子(RHLBP算子)获得的,该算子通过设定阈值能够更好地区分不同类别的纹理信息,且可以稳定表达相同类别的纹理信息。2.将颜色特征应用于面向对象的极化SAR图像分类。极化SAR数据并不能表示地物真实颜色信息,传统分类算法中,颜色信息未被用于提供分类依据,然而极化特征合成的伪彩色图像可以有效表达地物视觉信息。本文在分类算法中引入了伪彩色图像中的颜色特征,验证了颜色特征用于极化SAR图像分类的有效性。3.提出了一种新的基于原始数据的字典更新算法。用于稀疏表示的字典是从已知类别标签的数据中随机选取字典原子组成的,字典中很大可能会存在噪声点或者不能很好表示其他同类数据的字典原子。为了得到更有效的字典用于稀疏表示分类,本文提出了新的字典更新方法。该方法在原始字典中选取好的字典原子组成新字典,在保证字典大小相同的前提下可以获得更好的分类结果。4.提出了改进的联合稀疏表示算法。为了减少稀疏表示分类中对单个像素点分类产生的大量计算,本文引入了联合稀疏表示算法并对其进行了改进。原始联合稀疏表示算法将整体分割区域所有像素点用于求取共有模式,由于本文算法获得的分割区域大小不一致,且多数分割区域包含像素点过多,使得联合稀疏表示分类结果并不好。本文将分割区域内像素点分为数目固定的数据集进行联合稀疏表示分类,统计分类结果,根据统计情况对整体分割区域进行分类,可以获得更好的分类结果。
[Abstract]:Polarization SAR image classification as one of the important directions of remote sensing image processing and research has been paid more and more attention. Since the presence of a large number of noise points in polarized SAR data will seriously affect the further processing and research of the image, how to overcome the noise interference, extract the effective features and classify the polarimetric SAR image is a key problem in the polarimetric SAR image classification. In order to solve the above problems, this paper introduces the idea of object-oriented and sparse representation theory. The idea of object oriented classifies objects by initial segmentation to suppress the interference of a large number of noise points, and the robustness of sparse representation theory to noise also improves the correctness of classification. The combination of the two algorithms can obtain a good classification result. The four main contents of this paper are as follows: 1. The polarimetric SAR image is segmented initially. In this paper, the texture information is combined with the statistical region merging algorithm, and the Bhattacharyya distance merging criterion and the statistical region merging criterion are used to obtain more accurate image segmentation results. The extraction of texture information is obtained through the improved LBP operator, the locally consistent LBP operator (RHLBP operator), which can better distinguish different types of texture information by setting a threshold. And can stably express the same kind of texture information. 2. Color features are applied to object-oriented polarimetric SAR image classification. Polarimetric SAR data can not represent the real color information of ground objects. In the traditional classification algorithm, the color information is not used to provide classification basis. However, the pseudo-color images synthesized by polarization features can effectively express the visual information of ground objects. In this paper, the color feature of pseudo-color image is introduced into the classification algorithm, and the validity of the color feature used in polarimetric SAR image classification is verified. 3. A new dictionary updating algorithm based on raw data is proposed. The dictionary used for sparse representation is composed of dictionary atoms randomly selected from the data of known class labels. It is very likely that there will be noise points in the dictionary or that the dictionary atoms can not well represent other similar data. In order to obtain a more efficient dictionary for sparse representation classification, a new dictionary updating method is proposed in this paper. This method selects good dictionary atoms in the original dictionary to form a new dictionary, and better classification results can be obtained under the premise that the dictionary size is the same. An improved joint sparse representation algorithm is proposed. In order to reduce the amount of computation generated by single pixel classification in sparse representation classification, a joint sparse representation algorithm is introduced and improved. The original joint sparse representation algorithm uses all pixel points in the global segmentation region to obtain common patterns. Because the size of the segmentation region obtained by this algorithm is inconsistent and most of the segmented regions contain too many pixels, The result of joint sparse representation classification is not good. In this paper, the pixel points in the segmented region are divided into a fixed number of data sets for joint sparse representation classification, and the statistical classification results are used to classify the whole segmented region according to the statistics, and better classification results can be obtained.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52

【相似文献】