Binary data embedding framework for face recognition

In this paper, a supervised binary manifold embedding algorithm is proposed, which determines unique features for each class in a dataset. This algorithm is applied to a face recognition problem to find decision boundaries capable of separating the samples of the target face class from those of the remaining face classes. The proposed algorithm is realized in three different stages: Data pre-processing, Relation based weight generation and Embedding computation. The embedding computation stage applies the concepts of "friend closeness" and "enemy dispersion" in order to better control the relative positions of the important data samples from intraclass ("same") and interclass ("other") items. The method is a binary case of the Fisher criterion, and is employed for constructing the optimization templates for the proposed algorithm. The effectiveness of the proposed algorithm is compared with existing techniques using established datasets of faces - consisting of various poses, lighting conditions and facial expressions.