In this stage, our goal is to locate the centers of the dots in each
of the separated CMY channels, taking into account that these dots
were originally supposed to form a regular grid. Park et al. [2009]
describe a method for detecting 2D wallpaper patterns in real photographs
using mean-shift belief propagation. However, they face a
much more challenging lattice reconstruction problem and the performance
of their algorithm makes it impractical for recovery of
large screening grids. Liu [1996] recovers the halftone lattice parameters
in the Fourier domain, but his method does not seem to
account for grid distortions. Other parametric solutions encounter
the same difficulty. Thus, we propose our own simple and efficient
image-space approach. We start by detecting local maxima and
minima in each channel and then use least squares to recover the
grid explicitly accounting for the presence of a smooth deformation
field, which may arise in practice due to imperfections in both the
printing and the scanning processes.