Feature fusions for 2.5D face recognition in Random Maxout Extreme Learning Machine.

Abstract Contemporary face recognition system is often based on either 2D (texture) or 3D (texture + shape) face modality. An alternative modality that utilizes range (depth) facial images, namely 2.5D face recognition emerges. In this paper, we propose a 2.5D face descriptor that based on the Regional Covariance Matrix (RCM), a powerful means of feature fusion technique and a novel classifier dubbed Random Maxout Extreme Learning Machine (RMELM). The RCM of interest is constructed based on the Principal Component Analysis (PCA) filters responses of facial texture and/or range image, wherein the PCA filters are learned from a two-layer PCA network. The RMELM is an ELM variant where the activation function is based on the locally linear maxout function, in place of typical global non-linear functions in ELM. Since the RCM is a special case of symmetric positive definite matrix that resides on the Tensor manifold; a gap exists in between RCM and RMELM, which is a vector-based classifier. To bridge the gap, we flatten the manifold by transforming the RCM to a feature vector via a matrix logarithm operator. Experimental results from two public 3D face databases, FRGC v2.0 database and Gavab database, validated our proposed method is promising in 2.5D face recognition. Graphical abstract Highlights • A learning-based Regional Covariance Matrix (RCM) based on Principal Component Analysis (PCA) is proposed as a feature descriptor for 2.5D face recognition problem. • PCARCM is demonstrated as an intra-feature (range features) and inter-feature (range and texture features) fusion container. • Random Maxout Extreme Learning Machine as classifier is proposed to couple with PCARCM on the Tensor Manifold. [ABSTRACT FROM AUTHOR]