Research On Face Recognition Based On Deep Learning

Face recognition is an important research direction of biometrics.With the continuous efforts of many scholars and long-term exploration,face recognition has made a lot of achievements,but there is still a certain distance to solve the problem completely.In recent years,deep learning theory has developed rapidly,which provides some new ideas to solve face recognition problems completely.In this thesis,the author use the deep learning algorithm for face recognition and improve the face recognition effect further.The main contents of this paper include:(1)In this thesis,we first study the principle of principal component analysis,genetic algorithm and back propagation.In view of the problem that the face image data is large and easy to fall into the local optimal problem in BP,PCA algorithm and GA algorithm are used to optimize BP neural network and to form PCA-GA-BP network for face recognition.Firstly,the PCA algorithm is used to reduce the dimension of the face image and reduce the amount of face image data in the system.Then,we use the GA algorithm to optimize the weight of BP network.Finally,the AR database and ORL database are used to experiment.The results of the experiments show that the algorithm can not only reduce the amount of face image data,accelerate the convergence speed,but also reduce the probability that the BP network falls into local optimum and improve the recognition accuracy.(2)In view of the problem that the learning ability and non-linear mapping ability decline in a small range while the sample of BP neural network are increasing,deep belief network with super learning ability and non-linear mapping capability is proposed to replace the BP network,constituting the PCA-GA-DBNs network.The network uses GA algorithm and Gibbs sampling methods to carry out the network training layer by layer,and then use the BP network to fine-tune after training.Finally,the AR database and ORL database are used to experiment.The results of the experiments show that the algorithm can not only overcome the defects of BP network,but also improve the recognition effect of face recognition further.Meanwhile,the experiments also analyze the impact of different classifiers on face recognition.(3)In the case of large training samples,because of the lack climbing ability of GA algorithm,premature convergence appears easily.It is proposed that GA Algorithm should be replaced by SAGA,which has stronger global search ability and will not fall into the local optimum,forming the PCA-SAGA-DBNs network.The networks combine SAGA algorithm and Gibbs sampling to optimize the network layer by layer.After training,the network is fine-tuned and classifier is constructed by BP network.Then,the AR database and ORL database are used as subjects.The results of the experiments show that the network canovercome the shortcomings of GA's climbing ability and premature convergence easily and improve the recognition accuracy of face recognition at the same time.Finally,the three presented improved algorithms are tested.Results show that the PCA-SAGA-DBNs network not only has a good recognition effect but also has good stability,so it is a better face recognition method.