Fast and scalable enrollment for face identification based on Partial Least Squares

Face recognition has received increased attention due to its application in biometrics and surveillance systems, emerging two main tasks, verification and identification of faces. The first one aims at accepting or rejecting an identity assigned to a correct face, whereas the second aims at, given an unknown probe face, finding the best identity to it from a gallery of known faces. For the face identification problem, discriminative approaches such as the one-against-all method have achieved higher accuracy than descriptive approaches such as eigenfaces. However, such methods have scalability issues when new subjects are enrolled in the gallery once it is necessary to rebuild all discriminative models to take into account the new individuals. This work describes and evaluates a novel method for making the process of gallery maintenance more efficient. This method employs an association between the one-against-some classification scheme, which differently from the one-against-all approach that considers a random subset of subjects as counterexamples, and the use of a priority queue to provide a scalable approach to enrolling new subjects to the gallery. Experimental results obtained by applying the proposed method on publicly available face data sets demonstrate its advantage when compared to the one-against-all approach.