Abstract: In order to improve the performance of periocular recognition, a new method based on deep convolutional neural networks referred to as PeriocularNet is proposed. PeriocularNet exploited a 16-layer convolutional neural network, integrated with a residual learning module, and adopted the ArcFace loss function. Data augmentation was introduced to avoid the over-fitting in training process. The experiments on UBIPr and UBIRIS.V2 datasets show that the equal error rate (EER) of the proposed approach achieve 1.9% and 7.9% respectively. which improves the periocular recognition performance compared to the related methods. In addition, in order to verify the effect of the eyebrow region feature on the performance of periocular recognition in the end-to-end approach, two periocular datasets, UBIPr-1 and UBIRIS-1, involving three eyebrow shapes were established. Experimental results show that the EER of images containing the eyebrow feature is lower than that of the eyebrow feature removed, which indicates the importance of eyebrow feature in periocular recognition.