The Research Of Class Balanced Discriminant Analysis Feature Extraction Algorithm

Traditional Fisher Discrminant Analysis （FDA） does not consider the imformation for aspecific class, thus the recognition performance might be limited. The Class-Specific LinearDiscriminant Analysis （CSLDA） considers the imformation for a specific class，but might bringabout the problem of class imbalanace, which may have negative effect on the recognitionperformance. To solve this problem, this thesis borrows the idea of K-means clustering algorithem,and proposes three new feature extraction approaches.First, this thesis proposes an approach called K-means based Class Balanced DiscriminantAnalysis （K_CBDA）. For a specific class, K_CBDA first selects a reduced class, whose data arenearest to the data of the specific class, from the counterpart class, and further divides them intoseveral balanced subsets using K-means clustering algorithem. Then, K_CBDA performs LinearDiscriminant Analysis （LDA） on each sample subset, which is achieved by separately combiningeach balanced subset with the specific class, to extract discriminative vectors. Finally, K_CBDAuses a fusion scheme to construct a unified projection transform.Second, to further remove redundant information, this thesis proposes an approach calledK-means based Class Balanced Statistically Orthogonal Discriminant Analysis （K_CBSODA）.K_CBSODA introduceds statistically orthogonal constraint and makes the obtained projectivevectors statistically orthogonal. Furthermore, K_CBSODA decides the processing order of thesample subsets by considering the total scatter and discriminability, respectively.Third, to avoid the singular problem, this thesis borrows the idea of two-dimensionaldiscriminant analysis, and proposes an approach called K-means based Two-Dimensional ClassBalanced Statistically Orthogonal Discriminant Analysis （K₂D_CBSODA）.This thesis conducts experiments on the Coil₂0、USPS and Honda/USCD databases, and theexperimental results show that, the proposed approaches outperform several related methods.