Research On Binocular Vision Live Detection Method Based On Multi-feature Fusion Convolutional Neural Network

With the rapid advent of the mobile internet era,more and more systems use biometrics for security authentication.Face recognition has become the preferred biosafety authentication method in many situations because of its convenience,speed,cleanliness and other characteristics.Because face information can be reproduced in front of the face recognition system through personal photos and videos,the face recognition system has great security risks.In order to solve the security problem of the face recognition system,it is necessary to add the function of face live detection before the face recognition step to ensure the information security in the "big data" environment.To this end,carry out the following research work:(1)Propose a face living detection algorithm that uses an improved residual neural network to learn a homogenous binocular camera to produce a synchronized manual disparity map.With the continuous upgrading of camera equipment and display devices,the traditional live face detection methods have become invalid,and the existing binocular face live detection algorithms cannot completely solve the problem of face live detection.Since the real face is three-dimensional,the images viewed from different visual angles are different,and the face in the printed photo or playback video is flat,even if the bent photo is quite different from the real face.Secondly,the number of selfmade data sets in this scheme is relatively small,so the internal structure of the residual neural network is improved to increase the utilization of the internal features of the network.Experiments show that the homogeneous binocular face live detection of the improved residual neural network can effectively distinguish the attacking faces of printed photos and recorded videos.The homogeneous binocular face live detection algorithm is generalized compared to other binocular camera algorithms.The ability is stronger and the recognition accuracy is higher.(2)Propose a face live detection algorithm that combines color channel features and image gradient features and uses mathematical morphology structural elements instead of convolution kernel initialization.The generalization ability of a single feature is limited.A two-way convolutional face live detection algorithm model that combines gradient images and HSV color space is proposed;studies have found that some intermediate layers of neural networks randomly discard a certain proportion of neurons or weaken Certain feature channels of feature maps between the same level increase the sparsity between neurons at the same level,remove some redundant information,and speed up the network training speed and improve the model performance.In mathematical morphology,there are many "0" values in the structural elements of the morphological core,which can increase the sparsity of the neuron.Therefore,a multi-feature fusion live detection algorithm that uses the morphological core instead of the convolution kernel to initialize is proposed.This article involves the related content of binocular vision,convolutional neural network,and mathematical morphology.Experimental results show that this method can improve the accuracy of model recognition and speed up model convergence.