METHOD OF AND APPARATUS FOR SEGMENTING FOREGROUND AND BACKGROUND INFORMATION FOR OPTICAL CHARACTER RECOGNITION OF LABELS EMPLOYING SINGLE LAYER RECURRENT NEURAL NETWORK

摘要

A method and apparatus for processing a greyscale input of an image, particularly of a shipping label, into a binary output image in which foreground information is segmented from the background information and contrasts between adjacent regions having different background densities are obliterated is described. A neuron employing a 5 x 5 input neighborhood having a unique neuron activation function is shown. No explicit line process is employed. Output is biased toward a particular one of the output values by employing non-linear feedback as a function of both the grey scale value for the pixel corresponding to the label site being updated and the most recent value of the label site. The otherwise strong contribution from a gradient term in the energy function is suppressed by a shunting inhibition when the shunting inhibition function detects that the pixel lies on or near a boundary between adjacent regions of differing background intensities.