Deep representation alignment network for pose-invariant face recognition

With the recent developments in convolutional neural networks and the increasing amount of data, there has been great progress in face recognition. Nevertheless, unconstrained situations remain challenging, given their variations in illumination, expression, and pose. To handle such pose variation, we propose the deep representation alignment network (DRA-Net), which aligns the deep representation of the profile face with that of the frontal face. Comprised of a denoising autoencoder (DAE) and a deep representation transformation (DRT) block, DRA-Net uses end-to-end training. DAE recovers deep representations of large pose angle in not visible face areas, and the DRT block transforms the recovered deep representation from profile into near-frontal poses. Also, we implement cosine loss and use pairwise training to mitigate the gap between frontal and profile representations and reduce intra-class variation. In experimental results, DRA-Net outperforms other state-of-the-art methods, particularly for large pose angle on LFW, YTF, Multi-PIE, CFP, IJB-A, and M²FPA benchmarks.