基于均值漂移聚类的油罐识别.docVIP

基于均值漂移聚类的油罐识别 　　摘要：从高空间分辨率遥感影像中自动的提取油罐目标和具有重要的意义。针对传统方法计算复杂度高的不足，结合油罐和油库区域在遥感影像中体现的光谱和空间分布特点，提出一种基于均值漂移迭代聚类的油罐自动识别算法。算法基于均值漂移的分割算法进行区域的多层次合并，同时在每个分割尺度上依据形状参数检测油罐。多幅遥感影像的实验结果表明了算法具有较高的识别精度和的较低时间复杂度。 　　关键词：均值漂移聚类；自动识别；油罐识别 　　中图分类号：TP311文献标识码：A文章编号：1009-3044(2011)17-4155-03 　　A Mean Shift Clustering Method for Oil Tank Recognition 　　WANG Lei-guang1, ZHENG Chen2, DAI Qin-ling3 　　(1.School of Forestry, Southwest Forestry University, Kunming 650224, China;2.School of Mathematics and Statistics, Wuhan University, Wuhan 430072,China;3.Faculty of Wood Science Decoration Engineering, Southwest Forestry University, Kunming 650224, China) 　　Abstract: It has great significance for the military application to detect oil storage tanks and locate oil depots automatically from remote sensing images. Considering high computation complexity of traditional methods, an recognition algorithm based on mean shift clustering is presented, in which the spectral characteristics and spatial distribution of the oil tank is considering. Firstly, a multilevel hierarchical representation of image is established by using the mean shift segmentation approach to get segments in the finest scale and iteratively merging. At the same time, oil tanks in every scale are detected by evaluating shape features after every merging. The experimental results with several remote sensing images indicate that this method has high recognition precision, meanwhile, low time complexity. 　　Key words: mean shift clustering; convergence; automatic recognition; oil tank recognition 　　油罐等类圆形目标的识别是遥感影像军事目标识别的研究热点之一[1-2]。一般而言，霍夫变换（Hough Transform ,HT）是检测圆形目标最直观的选择[3-4]。然而，HT较高的时间和空间复杂度是其在具有海量特征的遥感影像处理中应用的瓶颈。考虑到油罐目标在高分辨率遥感影像中体现出较强的光谱同质性和分布密集型，可将油罐目标识别可视为聚类问题。油罐目标识别相当于空间位置上邻接、光谱同质性佳的像素的聚类问题。（Mean Shift，MS）算法是一种非参数的迭代聚类算法[5-6]，不需要事先指定聚类中心数目，也不需要对样本服从的分布做任何假设，可胜任任何特征空间的分析的场合。另外，由于迭代过程中样本点之间相互独立，可并行运算，因此适合海量遥感数据的处理。基于MS算法，本文提出了一种基于多层次区域合并的识别算法。首先，算法基于空间位置和灰度值特征的MS聚类[5]建立影像的多尺度表达；然后，通过在多个分割尺度上检测区域与圆形的接近程度，获得不同大小的油罐对应的区域