Binary to multi-level image restoration using neural network

An image processing method for converting N-level image data into M-level image data corresponding to the N-level image, comprising the steps of:

• determining coupling coefficients between neurons of a first neural network by repeatedly performing the following steps (a)-(e):
• (a) selecting a first object pixel of N-level image data, m-level (N<m<M) image data corresponding to which is known;
• (b) inputting the selected first object pixel and surrounding pixels of the first object pixel of the N-level image data to input neurons of the first neural network as input image data, each of the input neurons of the first neural network having been previously determined to input specific etude of the first object pixel and surrounding pixels;
• (c) inputting a second object pixel of the m-level image data corresponding to the first object pixel of the N-level image data as an ideal output for the input N-level image data to the first neural network;
• (d) comparing an output of the first neural network with the m-level image data input as the ideal output;
• and
• (e) correcting the coupling coefficients on the basis of a result of the comparison in step (d);
• determining coupling coefficients between neurons of a second neural network by repeatedly performing the following steps (f)-(j):
• (f) selecting a third object pixel of m-level image data, M-level image data a corresponding to which is known;
• (g) inputting the selected third object pixel and discrete surrounding pixels of the third object pixel of the m-level image data to input neurons of the second neural network as input image data, each of the input neurons of the second neural network having been previously determined to input specific data of the third object pixel and surrounding pixels;
• (h) inputting a fourth object pixel of the M-level image data corresponding to the third object pixel of the m-level image data as an ideal output for the input m-level image data to the second neural network;
• (i) comparing an output of the second neural network with the M-level image data input as the ideal output;
• and
• (j) correcting the coupling coefficients on the basis of a result of the comparison in step (i);
• successively selectively selecting each of the pixels of the N-level image data to be converted as a fifth object pixel;
• inputting the selected fifth object pixel and surrounding pixels of the fifth object pixel of the N-level image data to the respective input neurons of the first neural network;
• outputting image data from the first network as sixth pixel data of m-level image data corresponding to the fifth object pixel of N-level image data;
• inputting the sixth object pixel and discrete surrounding pixels of the sixth object pixel of the m-level image to the respective input neurons of the second neural network;
• and
• outputting image data from the second neural network as M-level image data corresponding to the fifth object pixel of N-level image data.

An image processing method for restoring input N-level image data to M-level image data corresponding to the N-level image, comprising the steps of:

• determining coupling coefficients between neurons of a first neural network by repeatedly performing the following steps (a)-(e):
• (a) selecting a first object pixel of N-level half-tone image data, M-level half-tone image data corresponding to which is known;
• (b) inputting the selected first object pixel and surrounding pixels of the first object pixel of the N-level half-tone image data to input neurons of the first neural network as input image data, each of the input neurons of the first neural network having been previously determined to input specific data of the first object pixel and surrounding pixels;
• (c) inputting a second object pixel of the M-level image data corresponding to the first object pixel of the N-level image data as an ideal output for the input N-level half-tone image data to the first neural network;
• (d) comparing an output of the first neural network with the M-level halftone image data input as the ideal output;
• and
• (e) correcting the coupling coefficients on the basis of a result of the comparison in step (d);
• determining coupling coefficients between neurons of a second neural network by repeatedly performing the following steps (f)-(j):
• (f) selecting a third object pixel of N-level character image data, M-level character image data corresponding to which is known;
• (g) inputting the selected third object pixel and discrete surrounding pixels of the third object pixel of the N-level character image data to input neurons of the second neural network as input image data, each of the input neurons of the second neural network having been previously determined to input specific data of the third object pixel and surrounding pixels;
• (h) inputting a fourth object pixel of the M-level character image data corresponding to the third object pixel of the N-level character image data as an ideal output for the input N-level image data to the second neural network;
• (i) comparing an output of the second neural network with the M-level character image data input as the ideal output;
• and
• (j) correcting the coupling coefficients on the basis of a result of the comparison in step (i);
• determining coupling coefficients between neurons of a third neural network by repeatedly performing the following steps (k)-(o):
• (k) selecting a fifth object pixel of N-level image data including half-tone image and character image data, M-level image data corresponding to which and areas of half-tone image and character image data are known;
• (1) inputting the selected fifth object pixel and surrounding pixels of the first object pixel of the N-level image data to input neurons of the third neural network as input image data, each of the input neurons of the third neural network having been previously determined to input specific data of the fifth object pixel and surrounding pixels;
• (m) inputting data specifying a type of the fifth object pixel of the N-level image data as an ideal output for the input N-level image data to the third neural network;
• (n) comparing an output of the third neural network with the ideal output;
• and
• (o) correcting the coupling coefficients on the basis of a result of the comparison in step (n);
• successively selecting a sixth object pixel of N-level image data to be restored;
• inputting the sixth object pixel and surrounding pixels of the sixth pixel of the N-level image data to the first, second and third neural networks, respectively;
• determining if the sixth object pixel is half-tone image data or character image data on the basis of output from the third neural network;
• outputting image data from the first neural network as M-level image data corresponding to the sixth object pixel of N-level image data if the sixth object pixel is determined to be half-tone image data and outputting image data from the second neural network as M-level image data corresponding to the sixth object pixel of N-level image data if the sixth object pixel is determined to be character image data.