Research And Implementation Of Face Recognition Based On Convolutional Neural Network

With the rise of computer vision and artificial intelligence research,its application in life has become more and more extensive.As one of the most popular methods of biometric recognition,face recognition has become one of the research hotspots in the field of computer vision and artificial intelligence.Deep learning is also the focus of research in recent years.Among them,convolutional neural networks can extract the deep abstract features of images because of the characteristics of local connection and weight sharing.Convolutional neural networks play an important role in deep learning.status.The traditional face recognition technology extracts the shallow features of the face for classification and recognition.The shallow features are easy to cause the loss of face information,and the generalization ability is insufficient.The face recognition based on convolutional neural network is compared with the traditional face recognition.Technology can extract deeper abstract features of faces and avoid some traditional face recognition problems.In this thesis,the artificial neural network and convolutional neural network are deeply studied.Based on the classical LeNet-5 model,the convolutional neural network face recognition model is constructed.This thesis mainly studies the following aspects:First,the basic theoretical knowledge of artificial neural network and convolutional neural network is collated.The neural network model and backpropagation algorithm are described.The local connection,weight sharing characteristics,activation function and convolution of convolutional neural networks are described.The basic structure of the neural network is described.Second,the objective factor illumination in face recognition has a non-negligible influence on the recognition rate.At the same time,it is not difficult to collect a large number of face images in the practical sense.In view of the above problems,this paper proposes a CNN face recognition method LECNN combining LBP and data expansion.The algorithm uses the LBP feature face that is not sensitive to illumination as the input of the CNN,and expands the LBP feature face by using the method of generating the anti-network.Simulation experiments verify that LECNN can effectively improve the accuracy of CNN face recognition.Third,the optimization algorithm is crucial for the training of convolutional neural networks,which is directly related to whether the training can be completed and whether the network can converge.This paper analyzes several optimization algorithms commonly used in convolutional neural networks,focuses on the most widely used Adam optimization algorithms,proposes an improvement based on Adam optimization algorithm,incorporates Nesterov momentum theory and improves the second-order estimation bias correction.The dynamic learning rate of the algorithm is improved,which ensures the monotonous decreasing characteristic and avoids the case where the learning rate is large and small.Simulation results show that the improved algorithm is better than the original algorithm.Fourthly,a real-time face recognition system based on convolutional neural network and a self-built face dataset are constructed,and the effectiveness of the improved points is verified.