Hopfield neural nets are used to optimize point-oriented binary computer-generated holograms (CGHs). It can be considered as a parallel and iterative 'halftoning' process in the spatial frequency domain. The results are comparable to other iterative methods but require shorter computation times. In this process, the generation of the CGH by FFT, binarization, and IFFT is viewed as a "black box" with inputs and outputs consisting of 512 arrays containing an object of size 64 . The neural-network optimization feeds back the Fourier transform of the reconstruction error to update the neuron states, which correspond to the samples of the continuous hologram. To reduce the error of the reconstruction, the input is allowed to deviate from the original array in different specified ways. For example, a previously reported approach using Projection Onto Constraint Sets (POCS) varied only the region of the input array outside the object, while we allow the entire array to be modified, thus providing more freedom in the optimization. The method may be applied either to magnitude-only or phase-only holograms. A modification of the parallel updating function is also reported. Different optimization options are compared. Use of a practical printing model requires optimization under assumed constraints to test the convergence properties of the algorithm.
Library of Congress Subject Headings
Holography--Data processing; Fourier transformations; Image processing--Digital techniques; Neural networks (Computer science)
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Li, Guo, "Neural network for optimization of binary computer-generated hologram with printing model" (1995). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus