Abstract:

An improved optimal discriminant plane(IODP) proposed by Zhao can only be used in the supervised pattern. Based on this point, a novel method is presented to extend IODP to the unsupervised pattern. On the basis of optimizing the fuzzy Fisher criterion to calculate the first optimal discriminant vector, the second optimal discriminant vector with the orthogonal constraint and the conjugated orthogonal constraint of the fuzzy totalclass scatter matrix can be figured out. These two vectors constitute the orthogonalconstraintbased unsupervised and uncorrelated optimal discriminant plane(OUUODP). With these, a novel unsupervised feature extraction method is obtained. The experimental results for the CMUPIE face database demonstrate that this method can extract the features which are conducive to classification and is superior to principal component analysis and independent component analysis when the betweenclass difference is big.

Key words: unsupervised pattern, feature dimension reduction, optimal discriminant plane, face recognition