Periocular recognition in the wild with learned label smoothing regularization

Periocular recognition has been gaining attention as one of the promising biometrics as it contains rich information of the ocular, skin and eyes color as well as eyebrow. Present researches of periocular recognition in the wild mainly are based on the convolutional neural networks that are equipped with standard cross-entropy loss. Label smoothing regularization (LSR) has been recognized as an effective regularization technique for generalization improvement. LSR optimizes the network based on a weighted combination of cross-entropy loss and KL divergence of uniform and network prediction distributions. In this paper, we extend LSR to Learned LSR (L2SR) by considering learned smoothen prediction distribution instead of predefined uniform distribution. L2SR outperforms LSR at reducing intra-class variation and, thus, improve the generalization. Extensive experiments on three periocular in the wild benchmarking datasets demonstrate the effectiveness and superiority of our method.