Multi-scale Multi-block Covariance Descriptor for a Compact Face Texture Representation: Application to Kinship Verification

Abstract

Division-based strategies for face representation are common methods for capturing local and global features and have proven to be effective and highly discriminative. However, most of these methods have been mainly considered for descriptors based only on one single type of features. In this chapter, we introduce an effective approach for face representation with application to kinship verification that relies on pyramid multi-level (PML) face representation, and which exploits second order statistics of several local texture features such as Local Binary Pattern (LBP), quaternionic local ranking binary pattern (QLRBP), gradients, and different color spaces. The proposed approach consists of two main components. First, we model the face image using a PML representation that seeks a multi-block-multi-scale representation where several local texture features are extracted from different blocks at each scale. Second, to achieve a global context information, we compute the covariance between local features characterizing each individual block in the PML representation. The resulting face descriptor has two interesting properties: (i) thanks to the PML representation, scales and face parts are explicitly encoded in the final descriptor without having to detect the facial landmarks, (ii) the covariance descriptor (second order statistics) encodes spatial features of any type allowing the fusion of several state-of-the art texture features. Experiments conducted on two challenging kinship databases provide results that outperform state-of-the-art Kinship verification algorithms.