Abstract: In the image analysis field，there have been some studies on association rules mining for texture feature from the affair data set that generates from the image neighborhood pixels，but association rules just keeping the frequent itemsets having maximized items leads the lose of information of the image.Extracting the feature from image based on frequent itemsets is proposed，this method first selects out the candidate frequent itemsets founded on the degree of frequency and the locality distributing，then constructs the feature set by defining the value of spatial control for every frequent itemset.The simulation is done with remote sensor image，because the EM clustering algorithm is dominated by its initializing，the experimentation specifies the different number of clusters for every feature set and determines the final clustering results by comparing the log likelihood.Experiment results show that frequent feature set can provide satisfactory expression for image.

Key words: frequent itemset, feature extraction, remote sensor image, Expectation Maximization（EM） clustering

摘要： 在图像分析领域，已有不少研究探讨了通过构建图像相邻像素之间的事务数据集，对图像纹理关联规则进行挖掘，但纹理关联规则仅存留最大项的频繁项集会使得很多信息丢失。为此提出了基于频繁项集的图像特征抽取方法，该方法首先基于项集的频繁度及空间分布筛选候选频繁项集，再定义每一个频繁项集的空间表达能力值构建特征集。在遥感图像上进行仿真测试，针对EM算法对初始设置比较敏感的特点，采用了对同一特征集指定不同聚类数目并比较对数似然值确定最终聚类结果的方法。实验结果表明，提出的频繁集对图像特征具有较好的表达。

关键词: 频繁集, 特征提取, 遥感图像, 期望最大（EM）聚类