Improvement of facial recognition with composite correlation filters designed with combinatorial optimization

The performance of composite correlation filters for pattern recognition depends upon the proper selection of training images. These images are commonly chosen based solely on the experience of the filter designer in an ad-hoc manner. As result, there is no guarantee that the best training images are chosen. In this work, we propose an iterative algorithm based on combinatorial optimization for the synthesis of composite correlation filters optimized for facial recognition. Given a data set of face images the algorithm finds the optimal combination of training images for the synthesis of a composite filter with the best performance in terms of quality metrics. Consequently, facial recognition with correlation filtering is substantially improved. Computer simulation results obtained with the proposed approach are presented and discussed in terms of facial recognition performance and classification efficiency.