IMD-Isomap for Data Visualization and Classification

In recent years, some nonlinear dimension reduction methods, named manifold learning, have been proposed and widely used in data visualization and pattern recognition. Of them, Isomap is a representative, which can project data from high-dimensional space into low-dimensional space with local structure preserved perfectly. However, Isomap suffers from the topological stability and is sensitive to noise. Moreover, it can only run in a batch mode, so cannot be directly used in pattern classification. In this paper, firstly, an improved Isomap based on image distance, namely IMD-Isomap, is proposed. Because spatial information of images is considered in image distance, as our experiments will show, IMD-Isomap outperforms Isomap for data visualization especially when noise is added. Then, combining IMD-Isomap and generalized regression neural network, which has a good ability for approximation, a classification method is proposed. Experimental results showed that our methods are robust to noise for data visualization or image classification when compared with KNN, Isomap or eigenface.