基于直方图的遥感图像相似性检索方法比较

Comparative Studies on Similarity Measures for Remote Sensing Image Retrieval Based on Histogram

相似性度量是用于研究多源数据之间相似程度的,是对空间数据进行模式识别的基础。通过单波段遥感图像的检索对两组直方图相似性检索方法进行了实验研究,即基于特征向量的相似性度量和基于概率的相似性度量。实验中发现第一组相似度量中有两种以往较少用于遥感图像检索的方法表现出色,它们分别是2χ统计距离和相似夹角余弦度量。第二组实验中,针对其中包含较明显的目标物体且背景较为单一的遥感图像(其直方图可看作混合高斯分布),在类别可分离判据的基础上,根据K-近邻法则提出了一种计算该类图像之间相似值的方法。实验结果表明基于K-近邻法则的计算方法行之有效。所得出的结论将对多源数据分析中相似性度量的理解与选择有积极意义。

Similarity measure is usually used to study the similar degree between muhisource data, which is the basis of pattern recognition on spatial data. In this paper two kinds of similarity measures are experimentally investigated through some remote sensing image retrievals, they are feature vector based measures and probabilistic measures, accordingly two groups experiments are designed to compare the measures for application to remote sensing image retrieval. From the experiment results we find that in the first group two measures seldom used in the literature perform well, they are χ^2 statistical distance measure and cosine of the angle measure. And in the second group experiments, for computing the similarity degree of two images with their histograms obeying mixture Gaussian distributions, we present a method on the basis of class separability measures according to the K-nearest neighbor rule. The experiment results show that the method has good performance. We believe that the results described in this paper will be of significance in applications to muhisource data analysis.