Local subspace method for pattern recognition

Abstract

Local Principal Components Analysis, i.e. Principal Component Analysis performed in data clusters, is discussed and a neural algorithm is developed. This algorithmic tool is used for pattern recognition. A decision function based on the subspace method is generalized by introducing normalization matrix F and affine coefficients α, β. Assuming that feature measurements are Gaussian in data clusters, it is shown that the new method is equivalent to maximum likelihood method. However, no explicit knowledge on probability distributions of feature vectors in classes, is required. For handwritten numerals the technique reaches the recognition rate of about 99%.