Research Of Face Recognition Based On Convolutional Neural Network

Convolutional neural network is a new network model,which combines depth learning technology and artificial neural network technology.It is widely used in the field of image recognition.The advantage is that the application system has a clearer hierarchical structure and a stronger sense of local area,and can effectively combine the classification and feature extraction process,which is conducive to training data.The particularity of convolutional neural networks is mainly embodied in the following two aspects:neurons are not all linked together,and neurons in the same layer can share the link weights.This will help reduce the number of network model complexity and weight,and maintain the stability of network structure.In this paper,a neural network model is built based on LeNet-5 structure,and the main research is as follows:First,design cascade relations,using human eyes to locate.The most important aspect of face recognition is face alignment,and face alignment is closely related to eye location.Therefore,we design a cascade relation roll machine neural network,which can detect the human eye without any constraints.After several comparative tests,it is proved that the neural network has stronger detection performance.Second,the experiment is focused on three convolution neural network models to verify that the improved method can improve the network performance.This paper mainly studies the following three convolutional neural network models,which are Yale,B face database,PUT face database and AR face database.Firstly,in each face database,the excitation function and sampling mode are combined in different ways.These combinations are configured in three network models to analyze image recognition capability;Then we can improve the performance of the network in three ways,such as increasing weight attenuation,using dropout technology and increasing momentum;Finally,the three methods are added to the network model to verify whether the network performance has been improved.Third,it is proved that the image recognition performance of multi bar convolution neural network is better than PCA+SVM.The two-dimensional Gabor filter is added to the structure of the convolutional neural network,and the convolution neural network is constructed by using 5 scales.In each scale,were compared in the following two ways,the first way is to directly use convolutional neural network performs classification operation;the second way is the first feature extraction using convolutional neural network,through the PCA reduction,then SVM performs classification operation.In addition,we have to integrate the five scales of convolutional neural networks into a multi column convolutional neural network.By changing the feature dimension and comparing the multi column convolution neural network and the PCA+SVM method,it has been proved by many experiments that the multi column convolution neural network has priority over the PCA+SVM method in the image recognition performance.