Large Margin Coupled Feature Learning for cross-modal face recognition

This paper presents a Large Margin Coupled Feature Learning (LMCFL) method for cross-modal face recognition, which recognizes persons from facial images captured from different modalities. Most previous cross-modal face recognition methods utilize hand-crafted feature descriptors for face representation, which require strong prior knowledge to engineer and cannot exploit data-adaptive characteristics in feature extraction. In this work, we propose a new LMCFL method to learn coupled face representation at the image pixel level by jointly utilizing the discriminative information of face images in each modality and the correlation information of face images from different modalities. Thus, LMCFL can maximize the margin between positive face pairs and negative face pairs in each modality, and maximize the correlation of face images from different modalities, where discriminative face features can be automatically learned in a discriminative and data-driven way. Our LMCFL is validated on two different cross-modal face recognition applications, and the experimental results demonstrate the effectiveness of our proposed approach.