Research On Deep Learning Based Face Reconginion

Face recognition technology is one of an important research field in the computer vision.More and more scholars and companies have participated in the research of face recognition.As a kind of biometric technology,face recognition technology has its unique advantages in how to ensure social information security.Although face recognition technology has made great progress,but at the same time,it also faces many problems that need to be solved.In a controlled environment,face recognition technology has reached a very high level,but if in a uncontrolled environment,such as illumination,pose and expression conditions,face recognition algorithm can not effectively extract facial feature information,resulting in the decline of recognition rate.Therefore,how to extract characteristics of human faces in the complex environment has become the focus of people's research on face recognition technology.In the study of the current face recognition technology and related papers,in-depth study of the current face recognition technology and the latest research results.The performance of the two-dimensional Gabor filter in different lighting conditions was studied.At the same time,aiming at the problem that the convolution neural network is easy to fail into the local optimal solution during the training process,improving the training method of network.The main work of this article is as follows:In the network structure based on convolution neural network,the characteristics of the Alex Net network are fully studied and its advantages are used in the transformation of lenet-5 network.the lenet-5 network is aimed at handwriting fonts,it is not appropriate to directly use face recognition.Therefore,in order to obtain the high performance network structure,the paper studies the Alex Net network,making full use of its advantages,and construct the new network by choose the propriety size of the convolution kernel size and the activation function and the Dropout technology.The experiment on AR face library shows that the face recognition system constructed by the modified convolutional neural network has better recognition performance and robustness than other algorithms.Aiming at the problem of vulnerable recognition rate in face recognition,such as illumination and attitude change,the presupposition Gabor nucleus of the convolution kernel is used to extract the facial features,enhancing the extraction of the effective information of facial features in the light environment.At the same time,when use the BP reverse propagation algorithm to train the convolution neural network in the training process,and the convolution neural network is easy to get into the problem of local optimal solution.The network training phase is fine-tuned to improve the recognition accuracy and generalization ability of the network.In the experiments,the proposed algorithm in ORL face database and extended Yale face database B experiments show that the algorithm in this paper under the condition of larger changes in illumination environment still can maintain good performance of significant advantages,while other algorithm recognition rate change is bigger,we can find that the algorithm has good robustness and generalization ability.