Classification of Incomplete Pattern Vectors Using Modified Discrminant Functions

The problem of classifying incomplete pattern vectors with the discriminant function classifier designed using the MSE criterion design approach is considered. A method for modifying the complete space parameter matrix of the discriminant functions is developed. The method allows the classifier to maintain MSE optimality for operating in any subspace of the pattern space. For full flexibility, only the inverse of the covariance matrix of the data in the Φ-space need be stored in addition to storing the parameters of the complete space discriminant functions.