Optimization of real-time Gabor-based face recognition

Many research works have proven the superiority of Gabor Wavelets as a feature descriptor for face classification as compared to other methods. However, most implementations used some commonly used values for Gabor parameters and we proposed that the performance of Gabor-based face recognition can be improved by proper optimization of Gabor parameters. Here in this paper we used Ant Colony Optimization to find the most optimal value for several Gabor parameters - the step frequency, maximum frequency and σ - that produced maximum classification accuracy. Additionally, due to high complexity of Gabor wavelets computation, we examine as well the feasibility of Gabor wavelets to be used in a real-time face recognition system. Using AR face dataset, we tested two classifiers namely Nearest Neighbor and ensembles of NN to determine which classifier would best suit a real-time implementation and delivers best performance. As a result we found that proper optimization would yield as high as 18% increase in face classification accuracy and it is actually feasible to use Gabor wavelets in a real-time system. We recommend using holistic implementation and NN method as classifier in order to achieve real-time performance for a Gabor-based face recognition system.