The Relationship Between Kernel and Classifier Fusion in Kernel-Based Multi-Modal Pattern Recognition: An Experimental Study

Two distinct principles of multi-modal kernel-based pattern recognition, kernel and classifier fusion, are demonstrated to share common underlying characteristics via the use of a novel kernel-based technique for combining modalities under fully general conditions, namely, the neutral-point method. This method presents a conservative kernel-based strategy for dealing with missing and disjoint training data in independent measurement modalities that can be theoretically shown to default to the sum rule classification scheme. Results of comparative experiments indicate that the neutral-point technique loses relatively little classification information with respect to coincident training data, and is in fact preferable for independent kernels produced by different physical modalities due to its better error-cancellation properties.