Unsupervised morphing by conjugate image processing

Abstract

Unsupervised image morphing is a very challenging task in machine learning, whereas it is relatively easy and common for the human brain. Even a simple deformation of an object such as a translation or scaling operation follows a complex curved trace in a high-dimensional image space. The learning of such a complex curved manifold needs human assistance or sufficiently dense intermediate data points. Presented is a novel transformation framework with which a curved manifold describing the deformation of images turns into a simple linear structure in the transformed space. Therefore, image morphing can be achieved by simple linear superposition using the conjugate image representation.