一种改进的ISOMAP算法在图像检索中的应用

An Improved Kernel ISOMAP Algorithm with Application to Image Retrieval

传统的核化ISOMAP(K-ISOMAP)算法对于多个分散类簇数据集的低维映射不能较好地表现数据集的内在拓扑结构。针对此缺点,本文将对基于ISOMAP的多类多流形算法(MCMM-ISOMAP)进行核化,提出核化的多类多流形ISOMAP算法(K-MCMM-ISOMAP),该算法不仅使得多类数据集在降维后保持较好的内在拓扑结构,而且具备了K-ISOMAP算法的泛化能力,可以将测试数据直接映射到低维空间。因此,该算法可以在多类图像数据集中实现图像检索的功能。实验结果表明该算法与K-ISOMAP相比更具有效性。

The conventional kernel ISOMAP algorithm （K--ISOMAP） can not work well in keeping the intrinsic topology of datasets from multi--class clusters datasets in the low--dimensional space. In order to avoid this shortcoming, a novel algorithm named kernel multi--c/ass multi--manifold ISOMAP （K-- MCMM--ISOMAP） is proposed in this paper, which is the kernel version of MCMM--ISOMAP. The new algorithm doesn＇t only keep the intrinsic topology of datasets in low--dimensional mapping space, but also has the generalization of K--ISOMAP. It can directly map the test data to low--dimensional space. Therefore it can be applied to the image retrieval system consisting of multi--class image dataset. The experimental results show that the new algorithm is more effective than the K--ISOMAP.