Multi-task Face Recognition Based On Convolutional Neural Network

As one of the most representative method in biometric recognition,face recognition technology has received extensive attentions and researchs.With the continuous development and improvement of technology,face recognition technology has been widely used in public safety,mobile payment and other fields.Convolutional neural network which is one of the most representative algorithm of deep learning has achieved many good results in face recognition.But most of the existed algorithms only discriminant the identity information,ignoring the association between identity information and other factors in the image.These associations can provide auxiliary informations to improve the effect of face recognition.The multi-task learning method can simultaneously learn the identity information and the auxiliary informations provided by other factors in the image to achieve the purpose of improving the recognition rate.Therefore,this paper combines convolutional neural network with multi-task learning,proposes a multi-task face recognition method based on convolutional neural network,and verifies it through a large number of experiments.The specific work is as follows.First,proposes a multi-scale feature fusion convolutional neural network structure.The features of high dimension in the convolutional neural network are more suitable for complex tasks such as face detection and face recognition,which are fused to obtain richer features while simplifying the network structure.Experimental results proved that the network structure achieves good results in all kinds of algorithms that only use a single network for training with the same order of magnitude of training images.Then combines multi-task learning with the multi-scale feature fusion convolutional neural network which proposed in this paper,and add side tasks such as illumination,pose,and occlusion.Multiple tasks are trained simultaneously and influence each other to improve the effect of face recognition.The loss function and network structure are modified so that multiple tasks can share the network and parameters,and the same network can be used for learning at the same time.Experiments on the CMU-PIE database prove that the effect of multi-task face recognition in a constrained environment is significantly better than that of single-task.Compared with the experiment of single recognition task,the recognition rate of both the main task and the side task is improved.Through multiple experiments,it is verified that assigning different loss function weights for each side task has different effects on the main task,and finally found a optimal loss function weight allocation scheme.A face database CASIA-CFP for multi-task face recognition in an unconstrained environment is constructed,and experiments on the database confirmed that the accuracy of multi-task face recognition in an unconstrained environment is still higher than that of single task face recognition.Finally,constructs a face recognition system in which the above mentioned multi-scale feature fusion convolutional neural network and multi-task face recognition algorithm are applied.