Independent component analysis (ICA) has been developed as a decorrelation technique for high-order moment of input signals. ICA has been frequently applied to blind signal separation. On the other hand, feature extraction and pattern recognition is also being focused as one of prominent applications of ICA. This paper studies patterns recognition using ICA feature vectors. The recognition rate is estimated for twenty-six alphabet patterns and the result is considered by examining the generated ICA bases. As a result, we conclude that recognition using ICA feature vectors outperforms recognition using PCA ones when noisy training data are used for training of ICA bases.
展开▼
