Phase singular points reduction by a layered complex-valued neural network in combination with constructive Fourier synthesis

Abstract

We propose a novel layered complex-valued neural network to reduce singular points (SP's) in phase images to obtain digital elevation maps (DEM's) through phase unwrapping. First we prepare a SP-free distorted image for a wrapped image data by constructive Fourier synthesis. We patch fractions of the SP-free image at the SP locations of the raw image and, then, feed it to estimation layer of the network as the initial image. The estimation layer interacts with raw-image layer with a complex-valued neurodynamics to yield a better estimation in which the SP number is reduced. The proposal reduces the calculation cost of unwrapping process and also increases the accuracy of the output DEM. © Springer-Verlag Berlin Heidelberg 2003.