Representation and Estimation of Tensor-Pairs

Abstract

Over the years, several powerful models have been developed to represent specific elementary signal patterns, e.g. locally linear and planar structures. However, in real world problems there is often a need for handling more than one elementary pattern simultaneously. The straightforward approach of adaptive model selection has proven to be difficult and fragile. At the core of this problem is the vicious intractable search space created by having to simultaneously select models and corresponding samples. This calls for higher order models where multiple patterns are represented as one more complex pattern. In this work, we illustrate the advantages of this approach on data that has bi-modal tensor-valued distributions.The method uses first and second order invariants as a representation, and an eigenvector based solution for recovering the elementary tensor components. We show that this method allows estimation of the two tensors that best represent a given tensor distribution. This distribution can for example be samples from a local neighborhood. A bi-modal distribution will produce the two tensors corresponding to the peaks of the distribution. In addition, numbers indicating the amount of samples belonging to each sub distribution are produced. We demonstrate the potential of the approach by processing a number of simple tensor image examples. The results clearly show that new valuable information regarding the local tensor structure is revealed.