Research And Implementation Of Face Recognition Based On Convolutional Neural Network

With the rise of artificial intelligence and the development of deep learning,face recognition based on convolutional neural network has become a research hotspot in the field of computer vision.The convolution neural network realizes end-to-end operation,and extracts image features automatically.However,the convolutional neural network is difficult to be well trained,and can not play an effective role in face recognition in the specific case where the hardware computing resources are limited.In order to improve the training efficiency and reduce the harsh hardware requirements of the original network,this paper analyzes the AlexNet and VGGNet respectively,and designs a reasonable network structure.The main contents of the full text include:(1)Summarizing the basic theoretical knowledge of convolutional neural networks,and introducing the Caffe framework.Taking LeNet as an example,the structure of the convolutional neural network and the principles of each level are introduced in detail.The processes of forward propagation and back propagation are deduced theoretically in the end.(2)Because of limited hardware computing resources,face recognition based on improved AlexNet and VGGNet network is proposed.The improved method for AlexNet is to replace the existing large convolution kernel with a small convolution kernel overlay module and reduce the number of fully connected layers.The improved method for VGGNet is to reduce the number of full connection layers,increase the BN layer,and fine-tune the structure use existing parameters.The improved structure is tested on LFW data set in the end.Comparative analysis shows that the improved structure reduces the number of parameters,reduces the hardware requirements,speeds up the training speed,improves the face recognition rate,and achieves the desired goal basically.(3)Building a face recognition system.A simple face recognition system is designed based on Caffe framework and pre-training model,which achieves real-time face recognition and static image verification.The non-training data set is used to test the system recognition rate under different conditions.Experiments show that the system software has goodrobustness,high recognition rate,meet the design requirements.