Research On Face Recognition Based On Deep Convolutional Neural Networks

With the rapid development of information technology,how to accurately identify and protect information security has become a key social problem to be solved urgently.Face recognition has the characteristics of intuition,non-contact,non-mandatory,simplicity and so on.It has become the most important method of identity authentication at this stage.Age recognition,gender recognition and expression recognition have become research hotspots in face recognition in recent years,and have broad application prospects in safe driving,clinical medicine,distance education and other fields.This paper focuses on facial age,gender and facial expression recognition.The main work is as follows:(1)In the research of face age and gender recognition,an improved depth convolutional neural network model is proposed for face images taken under natural conditions.It mainly focuses on three aspects:firstly,the convergence speed of depth convolutional neural networks are slow because of their many layers.In this paper,cascaded small-size convolutional core is used in the first layer of the network to reduce model parameters while ensuring the same field of receptivity,and Batch Normalization layer is used to normalize the output features,set a higher learning rate to improve the network learning speed,thus improving the convergence speed of the network.Secondly,in the study of deep convolutional neural networks,over-fitting is relatively common.Aiming at the problem of over-fitting,this paper replaces the first full-connection layer with small-size convolutional core and global average pooling layer,and sets a moderate dropout ratio to effectively reduce the risk of over-fitting.Thirdly,the illumination and background of the face image taken in natural scenes change greatly,which has a great impact on the accuracy of face recognition.Therefore,this paper uses data enhancement technology to train concentrators.The brightness and contrast of face image are processed to reduce the influence of light environment on recognition accuracy.The experimental results on the Adience dataset show that the convergence speed of the network in this paper is improved,the risk of over--fitting is further reduced,and the data enhancement technology effectively reduces the influence of light environment factors on the recognition accuracy,the network has higher recognition accuracy compared with the network proposed by LEVI and AlexNet.(2)In facial expression recognition,facial images with facial expression tags collected in a certain environment are used frequently,but most of the facial images on the network do not have facial expression tagging.Therefore,compared with age and gender recognition,facial expression datasets are relatively small,and it is difficult to fit large-scale deep neural networks.In order to solve this problem,four famous convolutional neural networks,AlexNet,VGGNet,Google LeNet and ResNet,are trained by using the"pre-training-fine-tuning"strategy.The aim is to optimize the initial value of network weight parameters through the pre-training process,so that the weight parameters are close to the ideal value before the network falls into over-fitting,so as to improve the recognition accuracy of the network.Then choose VGGNet-16 which has the best recognition effect after pre-training on the FER-2013 data set,and fine-tune the parameters on the expanded SFEW data set to achieve a higher recognition accuracy.Then the network pruning technology is used to compress the model parameters,and the global greedy pruning algorithm is used to compress the model parameters on the premise of ensuring the recognition accuracy,so as to avoid the over-fitting problem caused by the small SFEW training set and the excessive network parameters.After network pruning,the FLOP compression of VGGNet-16 is 35.1%,the parameter compression is 65.3%,and the recognition accuracy is 80.7%,which is much higher than the SFEW data set recognition benchmark value of 69.1%.