Centralized Large Margin Cosine Loss for Open-Set Deep Palmprint Recognition

As one promising branch of biometrics, palmprint recognition has received significant attention and made extraordinary progress in the past decades. The crucial step of palmprint recognition is to extract the discriminative features for the subsequent identification or verification task. However, neither the traditional hand-crafted descriptors nor the deep convolutional neural network (CNN) with the original softmax loss shows satisfactory generalization ability under open-set settings. In this paper, we proposed an end-to-end method for open-set palmprint recognition by applying CNN with a novel loss function, namely, centralized large margin cosine loss (C-LMCL). The modified loss function compels the feature vectors from different classes to uniformly and separately distribute in the hyper feature space. At the same time, it makes intra-class feature vectors compactly gather to their corresponding class centers. Consequently, such trained model has the ability to generalize across unseen subjects and different datasets. Finally, a lot of experiments are conducted on two public palmprint datasets- Tongji and PolyU datasets. In particular, all the evaluations are made under open-set protocols that are more complex and challenging compared to the previous close-set scenarios. The experimental results on the Tongji and PolyU datasets indicate the superiority of our algorithm over the state-of-the-art performance. It effectively confirmed the bright prospects of employing palmprint information in biometric authentication.