<title>Optimal binarization of input images for holographic neural networks</title>

Abstract

The optical implementation of neural networks using volume holograms for weighted interconnections requires stable phase relation between input channels. This is particularly important for images with variable illumination. One way to solve this problem is to use binary inputs. The simplest binarization is the direct quantization, but this method has a number of disadvantages. Error diffusion algorithm is more robust under variable illumination since it keeps the original image characteristics.Testing the abilities of error diffusion to process gray level images, we simulated an adaptive neuron unit that has been optically implemented for experiment [1,2]. It has been shown that the error diffusion is an effective way to binarize stationary gray level images for their utilization as input vectors of neural networks [3]. We show in the present work that it might provide advantages as better feature extraction of the common features and fast learning convergence of neural networks even under varying illumination ofinput images during the learning process.