PURPOSE: To improve the precision of identification by automatically generating data for identification which is weighted with the characteristic strength of plural graphics from picture information by means of updating learning for plural times.
CONSTITUTION: In data initialization for identification (1), the load values Kxxx of respective nets are defined as data for identification. A specified initial value or a value generated by random numbers are set. An initial graphic is set in recognition input as a learning graphic among plural graphics to which numbers are given (2). Furthermore, the load values Kxxx in which inputted recognition input is set to be identification data are used and the respective nets are operated, and respective recognition values are calculated in an output layer (3). Respective recognition values are compared and evaluated as to the recognition possibility of the learning graphic in such a state (4). When recognition cannot be executed as the result of evaluation, the load values of the respective nets are updated by a prescribed rule or the random numbers (5). Then, learning is continued and learning is terminated when the whole graphics can be identified without updating data for identification is updated once. (6).
TANAKA HIDEKI
AOYAMA JUNICHI
NA DETSUKUSU KK