Efficient boosting for synthesizing a minimally compact reduced complexity correlation filter bank for biometric identification

This paper addresses how to efficiently select which training images to use from an enrollment video sequence to train a correlation filter based face recognition system. Efficient enrollment and the selective use of training images from a video sequence of face images is a very vital component that determines the success of any face recognition system. We describe an efficient boosting algorithm for synthesizing a minimal set of filters that capture the different facial variations during the enrollment phase such that the resulting filter bank can also maintain good generalization and discrimination for recognition and verification. This is done by determining a fitness metric for each filter that determines the amount of facial variation capacity represented by that filter. If that capacity is exceeded by using more training images than needed for that filter then the resulting filter quality is compromised and discrimination performance can degrade leading to lower acceptance and rejection rates. We use advanced correlation filters that have shown to exhibit built-in illumination tolerance.