No abstract uploaded!

No abstract uploaded!

No abstract uploaded!

A collaborative convex framework for factoring a data matrix <i>X</i> into a nonnegative product <i>AS</i> , with a sparse coefficient matrix <i>S</i> , is proposed. We restrict the columns of the dictionary matrix <i>A</i> to coincide with certain columns of the data matrix <i>X</i> , thereby guaranteeing a physically meaningful dictionary and dimensionality reduction. We use <i>l</i> _1, regularization to select the dictionary from the data and show that this leads to an exact convex relaxation of <i>l</i> ₀ in the case of distinct noise-free data. We also show how to relax the restriction-to- <i>X</i> constraint by initializing an alternating minimization approach with the solution of the convex model, obtaining a dictionary close to but not necessarily in <i>X</i> . We focus on applications of the proposed framework to hyperspectral endmember and abundance identification and also show an application to blind source separation of nuclear magnetic resonance data.

Geometry processing algorithms often require the robust extraction of curvature information. We propose to achieve this with principal component analysis (PCA) of local neighborhoods, defined via spherical kernels centered on the given surface $\Phi$. Intersection of a kernel ball $B_r$ or its boundary sphere $S_r$ with the volume bounded by $\Phi$ leads to the so-called ball and sphere neighborhoods. Information obtained by PCA of these neighborhoods turns out to be more robust than PCA of the patch neighborhood $B_r \cap \Phi$ previously used. The relation of the quantities computed by PCA with the principal curvatures of $\Phi$ is revealed by an asymptotic analysis as the kernel radius $r$ tends to zero. This also allows us to define principal curvatures 'at scale $r$ in a way which is consistent with the classical setting. The advantages of the new approach are discussed in a comparison with results obtained by normal cycles and local fitting; whereas the former method somewhat lacks in robustness, the latter does not achieve a consistent behavior at features on coarse scales. As to applications, we address computing principal curves and feature extraction on multiple scales.