Neural net based digital halftoning of images

Various novel techniques for digital image halftoning are presented, performing nonstandard quantization subject to a fidelity criterion. Hopfield-type networks can be used for this task, minimizing a frequency-weighted mean squared error between the input (continuous-tone) and the output (bilevel) image. A novel kind of massively parallel analog network (the differential neural network) is introduced and shown to be appropriate for this task. This kind of network contains a nonmonotonic nonlinearity in lieu of the sigmoid function.<
>