Character recognition apparatus equipped with a recognition dictionary creation function

摘要

PURPOSE:To eliminate the need for executions of learning many times by updating respective values which are elements of a feature conversion matrix and a dictionary matrix through the use of a learning method with a back propagation. CONSTITUTION:In three-layer construction in a non-linear processing method, a matrix composed of coupling coefficients from a first layer 101 to a second layer 200 is regarded as a matrix for transforming a feature vector in a character recognition method, the matrix composed of the coupling coefficients from the second layer 200 to a third layer 300 is regarded as a recognition dictionary matrix, and the initial values of respective coupling coefficients are made into values obtained by the linear discrimination analysis of a learning vector. For the threshold of the second layer 200, '0' is made into the initial value, and for the threshold of the third layer 300, the threshold for 2-class identification is made into the initial value. Thereafter, the learning by the back propagation which is generally started with values given by random numbers can be started with optimum values in the case of linear formalism. Thus, a fast convergence, namely, a learning frequency is sufficient to be small.