Diffusion maps-based image clustering

In the clustering of large number of images using low-level features, one of the problems encountered is the high dimensional feature space. The high dimensionality of feature spaces leads to unnecessary cost in feature selection and also in the distance measurement during the clustering process. In this paper, we propose an approach to reduce the dimensionality of the feature space based on diffusion maps. In the proposed approach, each image is represented by a set of tiles. A visual keyword-image matrix is derived from classifying these tiles into a set of clusters and counting the occurrence of each cluster in each image of our database. The visual keyword-image matrix is similar to the term-document matrix in information retrieval. We use diffusion maps to reduce the dimensionality of visual keyword matrix. By reducing the dimensionality of the image representation, we can save computation cost significantly. We compare the performance between the proposed approach and the approach that uses the global MPEG-7 color descriptors. The results demonstrate the improvements.