Convolutional Neural Network Based Face Recognition Access Control System

Face information is used in various verification places mainly because it is unique,real-time and difficult to forge compared to other authentication means.Compared with some large places such as airports and railway stations,residential communities have smaller occupant capacity and more stable occupant information,which can give more full play to the advantages of face recognition technology.In this paper,we design a face recognition access control system based on convolutional neural network to achieve contactless and fast access control detection.In this paper,face detection,live detection,face feature extraction and comparison are mainly studied and researched.The face detection is achieved by using Kernel Correlation Filter(KCF)and Adaboost fusion method.The Histogram of Oriented Gradient(HOG)features and color histogram features are fused in the KCF target tracking algorithm to perform multi-scale tracking of the tracked target so that it can track and detect targets moving from far to near;then the Adaboost algorithm is used in the target tracking frame for face detection.The test proves that the face detection using KCF+Adaboost fusion method in the same video is 0.7s faster than the convergence speed of Adaboost algorithm alone.The fusion of Moiré texture and blink detection is used for live detection.Moirétexture detection is first performed on the detection target,and if it passes the detection,blink detection is performed.The blink detection is performed by first localizing the human eye region using the horizontal coordinates of 1 point,the vertical coordinates of3 points,and the horizontal and vertical coordinates of 13 points obtained by the Supervised Descent Method(SDM),followed by the detection and localization of the iris using the Hough circle detection method,and finally by comparing the iris region and the human eye region.Finally,the eye state is determined by comparing the ratio of the average grayscale of the iris region and the human eye region.The efficiency of live detection is improved by fusing the static moiré texture features with the dynamic blink features.A model based on a depth-separable convolutional CNN + Squeeze-and-Excitation(SE)module is constructed.The depth-separable convolution is used to reduce the number of operations and running time,and the network substructure of the attention mechanism of the SE module is used to enhance the learning of inter-channel correlation,strengthen the features of important channels and weaken the features of non-important channels,and finally train a convolutional neural network model with good accuracy.The face recognition access control system based on convolutional neural network designed based on the above method has been tested and proved to have an accuracy rate of 94.74%.