Deep Face Recognition Method Based On Sptio-Temporal Feature Fusion And Adaptive Loss Function

With the rapid development of computer technology and deep learning,image-based face recognition technology has been successfully applied in important fields such as smart payment and attendance systems.Such kind of methods heavily rely on the quality of the face image,however,the quality of the face image obtained in the actual scene is uneven,and direct recognition will cause the recognition accuracy to drop to varying degrees.Therefore,it is very necessary to study the face recognition technology with image sequence as input.This article mainly studies how to use reinforcement learning,deep learning and recurrent neural network to detect and recognize faces in image sequences.The specific work is as follows:(1)A method of face detection quality evaluation and classification based on deep reinforcement learning and self-adjusting reward mechanism is proposed.The existing face detection methods can only detect the position of the target's face while the function of assessing the quality of the detected face is absence.Our method divides the face detection into two stages.In the first stage,the Wider Face dataset is used to train the Retina Face network to detect the specific location of the face.In the second stage,reinforcement learning is applied to train a shallow neural network as an agent to classify the quality of the detected faces.Since the agent is not easy to converge in the initial training stage,this work designs a reward function that can automatically adjust reward to ensure the smooth training of the model.The effectiveness of this method is proved through comparative experiments on the CNN network with the same structure as the agent and the Face QNet face quality evaluation method.(2)A face feature extraction method based on deep learning and spatio-temporal feature fusion is proposed.This method uses the image sequence as input.First,the temporal feature extraction module Conv GRU is used to pre-extract the temporal features of the face image sequence;then use the method in Chapter 2 to select key frames from the image sequence and send them to the Resnet50 or Mobilenet backbone network to extract three different spatial feature maps,then send them to the spatial feature adaptive fusion module ASFF for weighted fusion,while the weighting coefficients are adaptively adjusted by the network;finally,the temporal features and spatial features are spliced together.Further fusion will be conducted to obtain more robust spatio-temporal features,thereby improving the performance of the face recognition model.Through ablation experiments,it is found that the accuracy of the method proposed in this work can be improved by about 1.2% compared with Resnet50 and Mobilenet networks.(3)A face recognition method based on deep learning and sample attention enhancement is proposed.Aiming at the problem that the AM-Softmax loss function used by the existing Arc Face face recognition network uses a fixed decision boundary and cannot fully adapt to the distribution characteristics of the training samples,this method proposes an adaptive decision boundary loss function based on the AM-Softmax loss.While retaining the original advantages of AM-Softmax,this loss function can adaptively enhance the attention to samples with large intra-class differences to automatically modify the decision boundary,thereby improving the convergence speed and recognition accuracy of the model.The ablation experiment proves the effectiveness of the method,and the comparative experiment on the MS1 M data set with the Cos Face face recognition network verifies the advancement of the method.