【作者】 冯静；

【导师】 焦李成；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 随着Internet技术的出现和通信技术的迅猛发展,人类积累的数据无论从规模上还是维数上都大大增加了,使得一些传统经典的聚类算法的效果急剧下降。为此,本文研究了投影寻踪线性数据降维方法,将免疫克隆算法用于投影寻踪聚类,对纹理图像、SAR图像进行了分割,取得的主要成果如下:(1)提出了基于免疫克隆的投影寻踪聚类算法,利用免疫克隆算法优化投影寻踪的投影指标,得到最优投影方向以及最优子空间,在最优子空间上获得了较好的聚类结果;(2)提出基于LDA投影指标的免疫克隆投影寻踪聚类算法,在投影寻踪聚类模型中,使用线性判别分析中的类间散度与类内散度的关系作为投影指标,并利用免疫克隆算法对这些指标进行优化,得到最优子空间。同时,对基于LDA投影指标投影寻踪聚类模型与K-means聚类算法进行自适应迭代优化结果,获得了较优的聚类结果;(3)根据拉普拉斯图建立有标签数据和无标签数据的关系,提出了基于半监督的投影寻踪聚类算法,并对纹理图像、SAR图像进行分割,实验结果表明,图像的分割精度有所提高;此外,将迁移学习引入基于LDA投影指标的免疫克隆投影寻踪聚类算法,对高光谱遥感图像波段迁移后进行图像分割。结果表明,图像分割精度有所提高。更多还原

【Abstract】 With the drastic development of Internet and communication, the accumulation of data is increased both in scale and dimension, which lead to the efficiency of some traditional clustering algorithms reduced. In this thesis, the dimensional reduction method of projection pursuit is researched, which is combined with immune clonal method to construct some new algorithms. Then these algorithms are applied in texture image and SAR image segmentation. The main achievements are as follows:(1)The algorithm of projection pursuit clustering based on immune clonal selection algorithm is proposed. Immune clonal selection algorithm is utilized to optimize the projection index, and then the optimum direction and subspace are obtained. Finally, the better clustering results are acquired in this subspace;(2) The algorithm of immune clonal selection projection pursuit clustering based on LDA index is proposed. In projection pursuit clustering algorithm model, we adopt Immune clonal selection algorithm to optimize the projection index which is the relationship between with-in scatters and between-in scatters, then the optimum direction and subspace are obtained. Meanwhile, the projection pursuit method and K-means clustering are iterated adaptively, and then the better results are obtained;(3) Based on the relationship of labeled and unlabeled data which is constructed by graph Laplacian, the algorithm of semi-supervised projection pursuit clustering is proposed. And the algorithm is applied in texture and SAR image segmentation. Seen from the results, the precision are improved. Moreover, transfer learning is introduced to the algorithm of immune clonal selection projection pursuit clustering based on LDA index. Firstly, transferring the band of the hyperspectral remote images, then do image segmentation. The result of segmentation shows higher precision.更多还原