Pattern recognition by polynomial canonical regression

摘要

A pattern is conceived as an equivalence class of samples in a criterion space. Two cases are distinguished: (1) The number of classes is very small in comparison to the number of samples; (2) it approximates the number of samples.Prediction of class membership by means of predictor variables in the first case is replaced by identification of a small vicinity of a sample in the second one.Linear and approximations of nonlinear canonical regression functions are proposed for both issues. Preprocessing of data by Karhunen-Loève expansion is discussed.Empirical results concerning classification of schematic faces and identification of photographic portraits are presented.