Face Recognition Based On Deep Residual Network

In recent years,with the rapid development of artificial intelligence technology,artificial intelligence technology has been applied in many fields.The problem of fast detection and recognition of targets by machines has always been a hot spot in the field of machine vision.Face recognition is the most unique natural way of identity recognition,and the advantages of face recognition compared with other recognition methods are convenient,fast,high security,easy to implement and install.At present,face recognition has been widely used in access control system,camera system,attendance system and financial system.Traditional face recognition algorithms are based on the distance between feature points or face subspace model.By comparing input images with trained features,face recognition can achieve high accuracy.However,the recognition accuracy is seriously affected by external interference,such as shading and back-lighting,facial pose angle,and masks with hats and eyes.Therefore,it is of great significance to find a recognition algorithm with higher recognition rate and stronger robustness.Based on the classical face recognition algorithm,this paper proposes a face recognition algorithm which combines the traditional feature extraction and deep learning.Because of the limitation of feature information extracted by Local Binary Pattern(LBP)operator and the problem of inaccurate description of image contour information.Therefore,Gradient Direction Histogram(HOG)and LBP hierarchical feature fusion are used to extract features from training set in convolution neural network,and then the matched extracted feature images are input into the improved convolution neural network for training and recognition.Because the original image contains a lot of noise and redundant information in the whole face recognition experiment process.In this paper,based on the previous part of experiments,a face recognition algorithm based on multi-directional feature extraction and composite deep residual network is proposed.The face image noise and redundant information are filtered through feature extraction process,and the algorithm is trained and recognized in the composite depth residual network composed of regression depth residual network and classification depth residualnetwork,so as to further verify the robustness and accuracy of the algorithm.The main contributions of this paper are as follows:(1)Feature extraction.The face image is encoded by binary code to extract LBP statistical histogram features.Then the histogram features of all the block regions are sequentially connected to form the feature matrix vectors of each layer to obtain the image.The HOG feature is extracted from the gray image again,and the hierarchical LBP feature vector is cascaded with the HOG feature vector in series to form the fusion feature of the face image.The fused features can provide feature data for in-depth learning training.(2)Deep Learning Network.On the premise of large data of face image,in order to improve the accuracy of face recognition and reduce the time complexity,the improved Convolutional Neural Network(CNN)and complex Deep Residual Network(ResNet)are used for deep learning.Deep learning network reduces training time and improves recognition rate through local perception field,weight sharing and residual block structure.