Face recognition: robust approach under varying and low resolution head poses

Abstract

In the last two decades, there have been many works in improving face recognition methods. Nevertheless, most of them are only reliable when strict conditions are applied which include small pose variations, constant illumination, and normal facial expression. In this paper we address the issue on performing face recognition across different face angles or poses. We proposed to treat the face features as vectors for both the target and the gallery faces and established angular relationship between these vectors. This angular relationship was first learnt through the training process. Using this angular relationship and with a given pose, the proposed method is able to estimate the feature vector of the full frontal face. Instead of considering the entire face, our proposed technique considers only local regions or face patches. Hence, given a face of different pose, the identity of the person could be recognized by relying only on single frontal face image. By removing the background region in some of the face patches, our proposed method performs considerably well virtually across all pose even for low-resolution face images.