Learning Discriminative Local Patterns with Unrestricted Structure for Face Recognition

Local binary patterns are a popular local texture feature for describing textures and objects. The standard method and many derivatives use a hand- crafted structure of point comparisons to encode the local texture to build the descriptors. In this paper we propose automatically learning a discriminative pattern structure from an extended pool of candidate pattern elements, without restricting the possible configurations. The learnt pattern structure may contain elements describing many different scales and gradient orientations that are not available in LBP (and related patterns), thus allowing the flexibility to construct structures capable of better representing the objects under test. We show through experimentation on two face recognition databases that this approach consistently outperforms other methods, in terms of training speed and recognition accuracy in every tested case.