Abstract
We propose the method to reduce the redundancy between features extracted from images by the infomax algorithm. The method consisted of the addition of noise and the feature selection based on kurtosis. The addition of noise to images contributes to the reduction of redundancy and make it easy to set a criterion of features selection. We adopt ETL-4 database sets of hand-written Japanese Hiragana characters, and performed the recognition experiment with the extracted features. Results of the recognition experiment showed that it was possible to reduce the redundancy of features by the proposed method.
