一种基于PCA的面向对象多尺度分割优化算法

多尺度分割是图像面向对象分类的基础，针对不同区域特征最优分割尺度确定的主观性以及采用聚类算法时聚类中心确定的随机性，提出了一种联合降维与聚类算法的面向对象多尺度分割优化算法。该算法首先利用主成分分析法（PCA）降维排序后的结果产生初始聚类中心；然后采用K-means聚类和度量每一个像素点合并的概率，从而得到适应不同研究区域内不同尺度地物的分割结果。采用多个影像数据库，通过引入聚类评价指标（内部评价指标和外部评价指标）、分割评价指标（分割精度、过分割率和欠分割率）并结合现有的图像分割方法及原始的K-means算法、与PCA降维后的K-means聚类对比分析。研究结果表明:经过降维处理后进行的聚类算法稳定性更高；与传统的聚类算法相比，结合PCA降维更能自动识别最优分割尺度；降维技术和聚类算法联合之中，目视和定量评价指标表明经过降维预处理后的聚类能得到更高质量的分割结果。 

An object-oriented multi-scale segmentation optimization algorithm based on PCA

Multi-scale segmentation is the basis of remote sensing images object-oriented classification. The paper proposes an object-oriented multi-scale segmentation optimization algorithm which combines dimension reduction technique with clustering algorithm aiming at the subjectivity of optimal segmentation scale determination of different regional features and the randomness of clustering center determined when using clustering algorithms. In this method, the initial clustering center is generated using the result of dimension reduction and sorting by Principal Component Analysis (PCA). Then the probability of merging each pixel is calculated by K-means clustering algorithm, so as to obtain the multi-scale segmentation results suitable for different scales in different research areas. This paper comparatively analyzes, in combination with the existing image segmentation methods and the original K-means algorithm, the K-means clustering segmentation after PCA dimension reduction, using multiple image databases, through a series of clustering evaluation indicators (internal and external evaluation indicators) and segmentation evaluation indicators (segmentation accuracy, over-segmentation rate and under-segmentation rate) to evaluate the result of different methods. The results show as follows: first, the method of the clustering algorithm after dimension reduction is more stable than the original clustering algorithm; second, compared with the traditional clustering algorithm, the PCA dimension reduction can identify the optimal segmentation scale more automatically; third, in the combination of dimension reduction technology and clustering algorithm, visual and quantitative evaluation indexes show that the clustering after dimension reduction preprocessing can get higher-quality segmentation results.