Face Recognition Based On Attention Mechanism

With the rapid development of convolutional neural network technology,face recognition has become a research hotspot in the field of pattern recognition and artificial intelligence.Traditional face recognition methods mainly rely on manual features of specific scenes for feature extraction,which have poor robustness and limited application scenarios.The face recognition method based on convolutional neural network directly extracts the face feature information with higher discrimination and more expressive ability from the face sample image through the multi-layer cascaded complex linear structure,thereby improving the accuracy of face recognition.The core of face recognition technology is how to extract strong distinguishing facial features.In view of this,this article focuses on the feature extraction network structure,and explores the conventional face recognition methods and lightweight face recognition methods based on the attention mechanism.The main research work is as follows:(1)A face recognition method based on hybrid attention mechanism and improved deep residual network is proposed.The feature extraction network structure of this method is obtained by reasonable improvement on the basis of ResNet-50,and a smaller volume is adopted for the input layer.The product core is used to ensure the resolution of the feature map;the improved pyramid residual block is used to build the main structure,and a hybrid attention mechanism is introduced in the residual block to combine the channel and spatial information between the feature maps to enhance the extraction The ability of key features of the face;Dropout and batch normalization operations are added to the output layer to improve the generalization ability and training efficiency of the model.The additive angular margin loss function Arcface is used for training to enhance the compactness and inter-class differences of face samples.Experiments show that the designed method has a higher recognition accuracy rate on the public data set.(2)In view of the limitations of the hardware conditions of mobile devices,the conventional face recognition algorithm model is limited by a large number of parameters and calculations and cannot be deployed.A lightweight face recognition method combined with an efficient channel attention mechanism is proposed.This method is improved on the basis of the MobileFaceNet network structure,combined with the efficient channel attention mechanism to further improve the ability of the network face image to exchange information across regions,and the channel information between the feature maps is combined to make the network more focused on people Extraction of key features of the face.At the same time,the adaptively scaling cosine loss function AdaCos is used for training,without manual adjustment of super parameters,which improves training efficiency and convergence speed.Experiments show that the designed lightweight face recognition method has a high recognition accuracy rate on the public data set.