A new algorithm for linear discriminant function and its applications to pattern recognition of nuclear explosion

In this paper, in the light of the criterion of minimum misclassified samples, a new kind of training algorithm for a linear discriminant function is proposed, based on the recognition results of all single features for training samples. There are two special construction algorithms, namely, the multi-dimensional direct method and the two-dimensional recursive method, designed for searching the optimal weight vector direction. In contrast to some traditional linear discriminant functions, the new training algorithm is hardly influenced by the typicality and the quantity of training samples. In addition, it needs less computation time. The experimental results for pattern recognition of a nuclear explosion show that the specific training algorithm is simple, practical and effective