A Mixture-of-Experts (MoE) Framework for Pose-Invariant Face Recognition via Local Landmark-Centered Feature Extraction

Most real-world applications in video surveillance and biometric authentication rely on robust face recognition systems capable of dealing with multiple variations of pose, illumination, and expression within the processed images. In this article, we are proposing a Mixture-of-Experts (MoE) framework together with local feature extraction (SIFT) centered around facial landmarks to address pose-invariant face recognition. For this purpose, the framework performs facial landmark detection with a twofold objective. First, head pose classification is conducted by processing a set of landmark locations detected in a face image to spot the visible landmarks. Second, these visible landmark locations are regarded as keypoints, and SIFT descriptors are extracted from them. These descriptors are utilized as inputs to the base learners comprising a MoE system, which is trained to compute the similarity between the subject identity it was trained for, and the unknown identity of the subject from the input image. This similarity is employed later to perform face recognition. We propose two models to be used independently as base learners. The first one is GMM, whereas the second one is a novel GMM-based model (Mahalannobis Similarity) introduced in this work. A performance comparable with state-of-the-art methods is obtained on images with pose angles between ±90° on the CMU-PIE, and Multi-PIE databases.