Anisotropic tensor field filtering

Abstract

Anisotropic filtering is based on a non-homogeneous smoothing technique that preserves regions with high structure. In the scalar case, several algorithms have been proposed to obtain this behavior, but it is not the case for higher order data field (e.g tensors). In this paper, we present a new non-iterative anisotropic method to regularize vector and higher order tensor fields. The driving application for this work is getting a general filtering technique. Our scheme is based on a multidimensional anisotropic adaptive filtering. The method first estimates the local structure of the field, and then, combines the responses of the tensor field to a set of steerable filter basis, weighted adaptively by the local structure. Promising results are presented from experiments on both synthetic and real second order tensor data.