Two-dimensional phase unwrapping using neural networks

Imaging systems that construct an image from phase information in received signals include synthetic aperture radar (SAR) and optical Doppler tomography (ODT) systems. A fundamental problem in the image formation is phase ambiguity, i.e., it is impossible to distinguish between phases that differ by 2π. Phase unwrapping in two dimensions essentially consists of detecting the pixel locations of the phase discontinuities, finding an ordering among the pixel locations for unwrapping the phase, and adding offsets of multiples of 2π. In this paper, we propose a new method for detecting phase discontinuities. The method is based on a supervised feedforward multilayer perceptron neural network. We train and test the neural network on simulated phase images formed in an ODT system. For the ODT phase images, the new method detects the correct unwrapping locations where some conventional methods fail. The key contribution of the paper is a one-pass pixel-parallel low-complexity method for detecting phase discontinuities