A Restricted Boltzmann Machine-Based Approach for Robust Dimensionality Reduction

Data dimensionality is an important issue to be adressed by pattern recognition systems. Despite of storage and processing, working with high-dimensional feature vectors also requires complex optimization methods. A proper selection of the most important features is essential and dimensionality reduction techniques can also be applied in order to avoid dealing with more information than needed. One of the most important analytical techniques for such task is Principal Component Analysis (PCA). In this work we propose a novel and more robust dimensionality reduction approach based on the Restricted Boltzmann Machines (RBMs), neural networks able to learn the probability distribution of the set of training samples, identifying the best features to discriminate them, for face spoofing detection. Results of the proposed approach show that the features learned and extracted by RBMs are more robust than the ones analytically obtained by PCA for differentiating between real and fake facial images.