Abstract
We propose a novel method of feature selection for similar-shaped character recognition using genetic algorithms (GA). The feature is assigned to the chromosome, and value of “1” or “0” is given to the chromosome; corresponding to features that are respectively used and unused for recognition. The proposed method selects only genes in recognition experiments with which the recognition rate of training samples exceeds the predetermined standard as a candidate of the parent gene and adopts a reduction ratio in the number of features used for recognition as the fitness value, and has a mechanism decreasing the number of chromosomes which take the value of “1” while changing generations. This mechanism enables the number of features to be decreased maintaining the recognition rate constantly. On the experiment for similar-shaped character recognition, the proposed method achieved a higher recognition rate and larger decrease of the number of features compared with Fisher's criterion. Moreover, the feature visualization that displays the pixel which is the source of the feature offered an experimental result that GA automatically acquires difference of character shape guessed that man is using for distinction. The effectiveness of the proposed method was confirmed by these.
