Divide And Conquer Based Discriminative Feature Extraction And Its Applications

Discriminant analysis is an effective feature extraction technique, which has been widely used for many real world applications, such as image classification and biometrics. However, in the scenario of big-data, due to the fact that the distributions of samples have large variations, the performance of traditional methods might be limited. To solve this problem, we borrow the idea of "Divide-and-Conquer", and propose a series of efficient feature extraction approaches.First, we propose an optimal subset-division based discrimination (OSDD) approach. The basic idea is that, we can divide the sample set into several subsets using K-means clustering, and then extract discriminative features from each subset. Then we construct a unified projection space. This approach could reduce the computational cost significantly. However, a key problem is how to decide a proper number of subsets. To tackle this problem, we design a generalized stability criterion to find the optimal number of subsets automatically.Second, to further enhance the discriminability of OSDD, we extend it to the nonlinear space, and propose an optimal sub set-division based kernel discrimination (OSKD) approach. OSKD divides the sample set into subsets by using kernel K-means in the nonlinear space, and extracts nonlinear discriminative features. We also show that our generalized stability criterion can also be used to determine the optimal number of subsets in the kernel space.Third, the idea of "Divide-and-Conquer" also motives us to study discriminant analysis from different point of views. For example, one common strategy is to simplify a multi-class problem to many two-class problems. Unlike existing methods, we only choose a part of adjacent classes of a given class to extract discriminative features. We propose a kernel adjacent-class discriminant analysis (KADA) approach. For each class, we select a part of its adjacent class to construct a larger class, and extract one discriminant vector. And we finally construct a whole projection space. Furthermore, we impose the uncorrelated constraints to KADA, and propose a kernel uncorrelated adjacent-class discriminant analysis (KUADA) approach.Experimental results on the AR, FERET, CAS-PEAL face databases and PolyU palmprint database show the effectiveness of our approaches. And our approaches have smaller computational burden than traditional methods.