Segmentation of character and natural image documents with neural network model for facsimile equipments

We propose the use of neural network for segmenting character, natural image, and screened halftone image documents. The model network has learned the characteristics of Laplacian histogram of each document. This network consists of three sub-networks and a decisional circuit. The structure of each sub-network is feed forward four layers structured neural network consisting of 10×10 input cells, 20 first hidden cells, five second hidden cells and two output cells. Those sub-network are the one dividing the documents of character and natural images, the one dividing the documents of natural images and screened halftone images, and the last one dividing the documents of character and screened halftone images. Input patterns to sub-networks composed of 10×10 data of optical density of pixels having 8 bits´ gray levels. The data were processed to have 10% saturation characteristics to the full scales for compressing scratch noises. Furthermore, Laplacian and logarithm operation were adopted to emphasize differences of luminosity distribution between character, natural images, and screened halftone images. An error back propagation method was also used as a learning rule for the sub-networks. As a result, the total network presented successfully more than 90% accuracy of discrimination for each types of document