Visualization of color-texture images using locally linear embedding

Abstract

Natural images, such as textiles and corals, are three-dimensional objects where variations in scaling, rotation, illumination, and perspective projections lead to drastic changes in the perceived image. These images are also multidimensional data and dimensionality reduction techniques are useful to extract features to describe properties of the images. Using linear reduction methods may discard important information on the global geometry of the data. In this study, we use a nonlinear dimensionality reduction technique known as Locally Linear Embedding (LLE) to derive meaningful features on natural color-texture images. LLE outputs significant global attributes from linear fitting of well-sampled data.
LLE was computed for color-texture images and coral images. It was found that color and intensity are meaningful attributes, which cluster the coral images in the embedding space.