Classification and segmentation of vector flow fields using a neural network

Abstract.

The main goal of this paper is to describe a neural algorithm for classification and segmentation of vector flow fields. We propose to use the coefficients of their projection into an appropriate linear space as a feature vector for classification. The projection onto a suitable set of basis vectors is computed by satisfying global optimization criteria. Once the whole flow field is partitioned into a large number of small patches, two processes are performed. In the former, each small patch is classified using the associated projection coefficients estimated by using a least-square-error (LSE) technique implemented on a neural network. In the latter, segmentation into larger homogeneous regions is performed using a region growing method. Two application contexts are considered: analysis of oriented textures and 3D motion. The Lie group theory is used to identify the basis vectors suitable for defining the vector space describing the patterns of interest. In particular, the projection onto the image plane of the 3D infinitesimal generators of the 3D Euclidean group have proved to provide an effective description for the considered vector flow fields.