Abstract: The classical ISOMAP（isometric feature mapping，ISOMAP） method developed on reconstruction principle may not be optimal from the classification viewpoint. Besides，it is prone to suffer from the noise and the range of the neighborhood. In order to resolve these problems, a novel method called KIMDISOMAP for dimensionality reduction was presented. Firstly, a modified image euclidean distance is proposed and used to find the suitable neighborhood. Then, direct linear discriminant analysis (Direct LDA) is used to replace multidimensional scaling (MDS). Compared with ISOMAP, the experiments on face recognition show that KIMDISOMAP enhances the ability of classification and extends the range of the neighborhood. In addition, the KIMDISOMAP obtains a better performance than other algorithms for images classification with small noise and geometrical deformation.