Large-Population Face Recognition (LPFR) Using Correlation Filters

Practical face recognition applications may involve hundreds and thousands of people. Such large-population face recognition (LPFR) can be a real challenge when the recognition system stores one template for each subject in the database and matches each test image with a large number of templates in real-time. Such a system should also be able to handle the situation where some subjects do not have a sufficient number of training images, and it should be flexible enough to add or remove subjects from the database. In this chapter, we address the LPFR problem and introduce the correlation pattern recognition (CPR)-based algorithms/systems for LPFR. The CPR, a subset of statistical pattern recognition, is based on selecting or creating a reference signal (e.g., correlation filters designed in the frequency domain from training images) and then determining the degree to which the objects under examination resemble the reference signal. We introduce class-dependence feature analysis (CFA), a general framework that applies kernel correlation filters (KCF) to effectively handle the LPFR problem. In place of the computationally demanding one template for each subject design method, we introduce a more computationally attractive approach to deal with a large number of classes via binary coding and error control coding (ECC). In this chapter, we focus on using advanced kernel correlation filters along with the ECC concept to accomplish LPFR. We include results on the Face Recognition Grand Challenge (FRGC) database.