Abstract: A generator in a back propagation type neural network having an input layer, an output layer and an intermediate layer coupled the input and output layers forms initial values for connection parameters. A first generator produces an initial value W10 of a weight coefficient of each connection parameter of the intermediate layer from in-class covariant data SW and inter-class co-variant data SB over data inputted to the input layer. The produced values are set into respective synapses of the intermediate layer as connection parameters.

Abstract: A method and apparatus for using a neural network to categorize patterns from pattern feature data derived from the patterns. The average of all the learning data of the neural network is subtracted from the pattern feature data, and the result is input to the input layer of the neural network. The neural network outputs a value for each category and the pattern is categorized based on these values. The neural network includes an intermediate layer whose bias is set to zero and which includes a sigmoid transfer function which is symmetric with respect to the origin.

Abstract: There are provided an image processing method and apparatus which can correctly extract desired image information from input image information and can recognize the extracted partial image information. An outline of a character or the like is traced by using an arbitrary point in the input image information as a start point. The start point is moved when the outline reaches the edge of the input image. The outline is again traced from the moved start point. The outline tracing and the movement of the start point are repeated until the character can be extracted by the outline tracing. When the number of times the start point has been moved exceeds a predetermined value, the input image information is inverted.