Pose-Invariant Face Recognition in Surveillance Scenarios Using Extreme Learning Machine Based Domain Adaptation

Abstract

Face Recognition (FR) under adversarial conditions has been a big challenge for researchers in the computer vision community. FR performance deteriorates in surveillance condition due to poor illumination, blur, noise, and pose variation in test samples (probe), when compared to training samples (gallery). Even recent deep learning methods fail to perform well in such conditions. This paper proposes a novel framework called PIFR-EDA (Pose-Invariant Face Recognition using Extreme learning machine based Domain Adaptation) that performs pose-invariant face recognition (PIFR) in cross-domain settings. It consists of two stages where the first stage performs face frontalization using a single unmodified 3D facial model and the second stage performs the task of robust domain adaptation by simultaneously learning a category transformation matrix and an ℓ1, 1 -regularized sparse extreme learning machine classifier. The proposed method outperforms state-of-the-art shallow and deep methods (in terms of rank-1 recognition rates) when experimented on three real-world face datasets captured using surveillance cameras.