Research On Deep Neural Network Recognition Method Of Periocular Image

Currently,the common biometric identification technology need to obtain complete and clear biometrics.For example,the face recognition performance has been greatly affected due to mask occlusion and glasses.Therefore,biometric recognition technology under less-constrained scenarios has gradually become a research hotspot.The periocular region contains rich color and texture features.Compared with the whole face region,it is less affected by age change and expression change,and has higher differentiation and stability.Multi-mode fusion recognition such as face and iris or iris and periocular can combine the advantages of multiple biological modes and make up for the deficiency of single mode recognition.Therefore,based on the convolutional neural network,in this paper we conduct in-depth research on three aspects: periocular preprocessing,periocular recognition,and periocular and iris fusion recognition.The results show that periocular recognition and fusing recognition between periocular and iris are superior in less constrained scenarios.The main work and contributions of this paper are as follows:(1)A lightweight neural network called MEL-YOLO is proposed to preprocess periocular images.MEL-YOLO can simultaneously detect the region of interest of the human eye,identify multiple attributes of the human eye,and locate key points.The network can solve the problems that the performance of human eye localization is degraded under the influence of various interference factors,e.g.illumination,glasses and occlusion.MEL-YOLO combine YOLOV3with the enhanced DS-sandglass block,and propose a denormalized coding and encoding method in the key point regression to promote the network positioning depth.In addition,the complete intersection-over-union(CIo U)and the mean square error(MSE)are introduced in the paper as the loss function to promote the overall performance of the network.On the near-infrared dataset,MEL-YOLO achieves accuracy of 100%in human eye detection,and achieves98.7%and 96.5%in the attribute recognition and landmark positioning,respectively.On dataset UBIRIS,MEL-YOLO reaches 92%and 91%.The experimental results demonstrate that MEL-YOLO can simultaneously perform human eye detection,attributes recognition and key point positioning.Meanwhile,it is proved that MEL-YOLO is lightweight,robust,and has good generalization ability which can be applied to low-performance edge computing devices.MEL-YOLO lay a data foundation for subsequent periocular recognition and multi-modal recognition.(2)Periocular feature extraction approach based on deep convolutional neural networks is proposed.The powerful ability of convolutional neural network can extract more robust and accurate periocular features.In this paper,we design a new architecture called Periocular Code Net for periocular recognition tasks,which combines the improved DS-Sandglass module with the Cos Face Loss function that extends the decision boundary.In order to improve the performance of periocular recognition,we have done many experiments on the parameter m value of Cos Face Loss,the input form of periocular images,etc.The result on multiple dataset shows the proposed Periocular Code Net outperform both conventional and SOTA deep learning algorithms in Equal Error Rate(EER)and True Accept Rate(TAR),demonstrating that Periocular Code Net has a superior periocular feature extraction capability.(3)An algorithm for multi-modal fusion recognition of periocular and iris is proposed.Considering the superiority of multi-mode,a multi-modal(including periocular and iris)fusion recognition of study was conducted.The performances of two single modes and multi-modal were tested on CASIA-Iris-Lamp.The experimental results show that the multi-modal(including periocular and iris)recognition performance is better than that of any single mode,which shows the broad research prospect of multi-modal recognition technology.