Data dimension reduction in training strategy for face recognition system

In this paper, we propose a training strategy for an automatic face recognition system. Our strategy is based on cascade reduction of data dimensionality using LBP and PCA algorithms. This method is able to achieve higher recognition accuracy in comparison with simple LBP or PCA and it is suitable in the case of adding a new user to the face recognition system. We provide a comparative study of our proposed algorithm and several standard algorithms that reduce dimensionality of input data. Dimension reduction is important also in the case of storage and computational complexity reduction. We propose an overview of selected strategies and we compare their performance using CMU PIE face database. Our results in testing of clustering algorithms indicate that SOM and K-means algorithms are suitable for an automatic selection of training samples for a recognition system. According to achieved results we propose a part of the face recognition system suitable for example for the next-generation of hybrid broadcast broadband television.