The Real-time Video Face Recognition Research Based On The Deep Learning

In recent years,we can make use of face detection to get the images of human faces in the real-time video and recognize faces.The face recognition application get more and more people's attention.There is a variety of face recognition techniques,including shadow features extraction for recognition,deep structure for learning,etc,and these techniques have many different characteristics.Deep network model can be constructed by deep learning methods and represent essential and effective information of the complex images.So the use of deep learning algorithm to model the face image and the use of the constructed deep network model to identify face will have a very important significance.Deep Convolutional Neural Network is a classical and frequently-used deep structure.Deep Convolutional Neural Network is a effective way of learning in 2D face image recognition scene and is used to solve the problem of face recognition in extensive literature.However,at present Traditional pool sampling method that is used by Deep Convolutional Neural Network is single and fixed,and this sampling strategy will increase the likelihood that effective information is lost in layer and layer.Therefore,traditional pool sampling method reduces recognition performance and generalization ability of the network.In addition,the constructed convolution make use of established the whole connection methods in adjacent layers and can not effectively reduce the number of parameters to be trained.In order to solve these problems,an improved Deep Convolutional Neural Network is proposed.In the deep of the convolutional neural network,we first propose the sum of the square probability pool sampling method,and the new pool sampling method is based o n the square of element activation value in the pool area specific sampling as the basis for the calculation of probability.Then its activation value of each element multiplying the corresponding probabilities will be summed to obtain the sample pool of values sampled within the unit,so as to solve the traditional algorithm preclude the samples over a single and fixed the problem.Then,it is proposed to use random pattern on one half of the feature to get map features of the next level between the pool sampling layer and the convolution layer,that improve some connected way,so that the connection laminate more efficient and comprehensive between the pool sampling layer and the convolution layer,while reducing the number of parameters of the network need to be trained.Finally,for the design of high-rise structure,we will connect multiple levels of feature map with a full connection layer and Softmax regression layer and make supervised training in order to get the correct classification.Experimental results show that,compared to the former algorithm improvements and other classic face recognition algorithm,an improved deep convolution neural network structure that is proposed reflects better performance for face recognition and has more good theoretical significance and application value.