The Research On Face Recognition Based On Deep Learning

This paper sets identity verification as the research background.By analyzing the current research situation and deficiencies in the field of face recognition,combining with the deep learning methods,it puts forward a convolutional neural network model,which is aimed to improve residual modules and sliding windows,and study the face recognition accuracy as well as the computational efficiency of the model algorithm.Firstly,it introduces the research methods of face recognition at home and abroad,compares the advantages and disadvantages of feature extraction methods,and then proposes a method by using multi-layer convolutional neural networks to solve feature extraction in face recognition.Compared with the traditional face recognition method,the convolutional neural network model can conduct a more robust face representation by extracting different layer-by-layer features.In order to improve the accuracy of facial feature representation,this paper improves the face image segmentation method,which is based on the Deep ID model.By applying the sliding window cutting method,it makes the facial feature representation information more accurate.In order to reduce the computational cost,this paper will share a set of hidden layers of the convolution neural network model,and then segment the sliding window on the feature graph.In this way,the features of multiple regions in face can be extracted more efficiently.In order to extract higher order face image information,the depth learning model is getting deeper and deeper,but it leads to the problem that the performance of the network model is degraded,that is,the deeper the neural network is,the more difficult it is to train the neural network.This article refers to the network structure of Res Net,which solves the problem of degraded model computing performance by introducing the residual model.Due to the complex structures of the neural network and the large number of parameters,it uesd to cause an overfitting phenomenon in the training process.In order to solve the problem of excessive computation caused by the fusion of multiple features,this paper improves the convolution layer in the original residual module.It also introduces a 1 × 1 micro-convolution structure,which greatly reduces the number ofparameters and improves the model's training speed and saves the training time.Finally,this paper performs multiple comparison experiments on LFW and other human face databases,and obtained the optimal method for dimension reduction and similarity determination.By applying the method in the model,it effectively works in LFW people and achieves a 98.75% face recognition accuracy.