Information fusion by combining multiple features and classifiers

Information fusion by combining multiple sensor data and classifiers has received considerable interest in recent years. It is believed to be an effective way to design pattern recognition systems with high recognition accuracy, reliability, and robustness, to deal with the variance-bias dilemma, to handle different sensor or feature types and scales, and to maximally exploit the discriminant power of individual features and classifiers (experts). This paper provides a brief survey and a taxonomy of various schemes, which are characterized by their (i) architecture, (ii) selection and training of individual classifiers, and (iii) combiner's characteristics. Recent work on theoretical analysis on combination schemes is also briefly summarized.