A study on handwritten digits recognition using independent components

An application of independent component analysis (ICA) to characters recognition is proposed in this paper. The purpose is to evaluate effectiveness of features extracted by ICA. We propose a novel recognition system that consists of modules for each category. A module has two parts: a feature extraction and a classification. Features are independent components estimated by ICA and outputs of a classification using features are candidates for categories. These candidates are combined based on majority rule and categories are decided for input images. Hand-written digits in MNIST database are used as target characters. FastICA algorithm is applied to these images in order to learn a separating matrix. In recognition experiments, we demonstrated that ICA extracted useful features for handwritten digits and independent components were superior to principal components for the recognition accuracy. Furthermore, we showed the addition of noise pattern to training data was effective for elimination of redundant basis functions. From these results, we confirmed the effectiveness of the feature extraction using ICA.