The Research Of Facial Information Recognition Algorithm Based On Deep Learning

Nowadays,facial information recognition is widely used in all aspects of life.Face information recognition refers to detecting the face in a picture and analyzing the information contained in the face,such as the name,gender,age,race,gesture and expression of the face.Face detection is the basis of facial information recognition.Face recognition can confirm a person's identity and facilitate the recording of facial information.Facial expression recognition can identify the user's mood,mental state and other conditions,has a huge application scenario in medical,security,human-computer interaction.This paper focuses on the three aspects of face detection,face recognition,and facial expression recognition in facial information recognition of human faces.Due to the superiority of deep learning in the field of images,more and more scholars are using it to solve the problem of face detection.The current face detection methods based on deep learning can be divided into three main types: one type is One Stage detection method,which directly extracts features through convolutional neural networks and then classifies and regresses;One type is Two Stage detection method,which first extracts a large number of candidate regions through candidate region selection algorithm,then extracts features through convolutional neural networks and classifies and regresses the extracted features;And another type of detection method is constructs multiple weakly classified convolutional neural networks and cascades multiple networks to achieve face detection.In this paper,we study each of the three types of networks and finally select a cascade face detection method.The method in this paper is based on MTCNN and improved to implement a cascade-based face detection method.Training was performed using the Wider Face public face dataset,performance was evaluated on the FDDB dataset,and comparison of detection speed,and the proposed method was found to achieve better test results.In face recognition,most of the methods for deep learning suffer from problems such as deep network layers,large number of parameters and slow training.In this paper,the problem of excessive number of parameters and slow training speed is improved to some extent by improving the face residual network and replacing the conventional convolution in the residual network with depthwise separable convolution.At the same time,the model is optimized by improving the loss function into a triplet loss function to address the problem of the inadequacy o f the separability of the Softmax loss function.Through experiments,it was found that the triplet loss makes the intra-class spacing smaller and the inter-class spacing larger,so that facial features can be better recognized;the depthwise separable convolution reduces the network parameters by about 45% with little change in recognition rate.It was trained on the CASIA-WebFace face dataset and tested on the LFW face detection dataset with an accuracy of 99.36%.For facial expression recognition,the three network models of AlexNet,VGGNet,and ResNet are selected in this paper.The above three network models for deep learning are improved accordingly for the need of facial expression recognition.Considering the public facial expression data set,there are usually problems such as small data volume and uneven data distribution,so data alignment and data enhancement are performed during data preprocessing.Different hyperparameters are used for different network models during network training.ResNet's network model uses transfer learning to assist in training.Through training and testing on the Fer2013 and CK+ expression datasets,and comparing the accuracy and analysis confusion matrix,the improved VGGNet expression recognition network was selected as the facial expression recognition model.Accuracy rates of 99.31% and 70.43% were obtained on the Fer2013 and CK+ data sets.In practical applications,it is not enough to detect and identify information about a face,and it is necessary to build a synchronous recognition system for multiple information of the face.By realizing multimodal and synchronous recognition of facial information,facial information recognition can be applied in different scenarios and has higher use value.In this paper,we construct a facial information synchronization recognition system by combining face detection,face recognition and facial expression recognition algorithms.