Research And Application Of Key Techniques For Face Authentication Based On Residual Separation Convolutional Neural Networks

With rapid development of deep learning and neural networks,many major breakthroughs have been achieved in the field of computer vision.Face authentication is an indispensable part of computer vision with a wide range of application value.How to efficiently and accurately realize face recognition and authentication by computer technologies becomes the current frontier research topic.The two main steps of face authentication are: 1,Detect the face in the image to obtain the face detection frame.2,Characterize the detected face by feature extraction.The implementation of these two major steps would not be possible without convolutional neural network.This thesis designs a new convolutional neural network structure named residualseparation convolution block based on the residual structure and the separate convolutional models.It first explores the design philosophy of the residual-separation convolution block,and analyzes it in the ImageNet data set to verify its feasibility.Then the face detection network and the face recognition network are designed respectively using the residual-separation convolution block and a thorough comparison with other convolutional neural networks is conducted.Meanwhile,a simple face liveness detection classifier is also implemented using the residual-separation convolution block.Combined with Internet-related technologies,a network face authentication login system is implemented.The main content of this thesis includes:1.Combining the residual network and the deconvolution model in modern convolutional neural networks,the residual separation convolutional neural network is designed and implemented.Experiments with ResNet and MobileNet in the ImageNet dataset demonstrate the effectiveness of the residual separation convolutional neural network in image recognition speed and accuracy.2.Designing a multi-task cascaded face detection network with residual separation convolution block,which is 1.7% more accurate than the most widely used MTCNN face detection network.3.Designing the face representation network using the residual-separation convolution block.The network achieves 99.67% accuracy in the LFW verification data set with a 52-layer structure.At the same time,the test duration was 38 seconds and 18 seconds faster than ResNet and MobileNet,respectively,under the same experimental conditions.Also,the network achieved 72.41% accuracy in the unwashed million face challenge.Accuracy of 81.5% was achieved in the millions of face challenges after cleaning,and 98.97% accuracy was achieved when the number of distracters was reduced to 10,000.4.Designing live detection networks using residual separation convolution blocks.In combination with face detection network and face representation network,a web face authentication system was implemented using Internet-related technologies.The system has three characteristics: high availability,high agility,and high security.Tests have shown that this system is adequate for the various needs of web face authentication login.