Study On Convolutional Neural Network Expression Recognition Algorithm Based On Transfer Learning

Facial expression is an important information used to express and recognize emotions.With the rapid development of computer technology and artificial intelligence,many researchers devote themselves to the research of facial expression recognition in the field of computer vision.At present,the use of facial expression recognition with computer has been applied into many places of people's lives,such as intelligent teaching,interrogation investigation,digital entertainment,driver fatigue detection and so on,but it still has broad application prospects,so continuous exploration of facial expression recognition methods is of great significance to social development.At present,most of the research tasks of facial expression recognition are completed by convolution neural network.However,in order to improve the performance of facial expression recognition tasks,more and more levels and parameters have been found in convolutional neural networks.So the running speed is affected and the hardware devices is under pressure.It is important that deepening the network level cannot achieve the effect of extracting the expression features of the key areas completely and accurately.Another disadvantage of convolutional neural network is that it depends on a large amount of data to drive.At the same time,the data needs to be labeled and come from the same distribution.In the actual scene,it is difficult to obtain a large number of labeled expression data sets,so it has an impact on the use of convolutional neural network.Based on the theory of image processing and deep learning,this paper focuses on the accurate recognition of facial expressions,and focuses on the application of convolutional neural networks in facial expression recognition tasks,and realizes the effective recognition of facial expressions.The main research contents of this paper are summarized as follows:(1)Aiming at the problems of large convolutional neural network models,too many parameters,easy overfitting of models,and insufficient accuracy of feature extraction in key areas of facial expression images,a convolutional neural network facial expression recognition method based on attention mechanism is designed.In this method,a lightweight network is designed,and residual identity blocks are added to its convolution layer connection to alleviate the problem of model overfitting.At the same time,the attention mechanism spatial group wise enhance(SGE)module is introduced.Feature grouping and spatial enhancement are used to solve the problem of imprecise feature extraction in key regions of facial expression images.Then SGE uses the similarity of global features and local features to guide the spatial distribution of semantic features,enabling each feature group to autonomously enhance the feature learning of facial expressions.The amount of parameters introduced by the SGE module is negligible and has no effect on the training speed of the network.The comparison of each group of experiments shows the superiority of the proposed network expression recognition performance.(2)In view of the current convolutional neural network training that needs to rely on a large number of labeled data drives,the training and test data samples must come from the same feature space and the same distribution,and at the same time,in the facial expression recognition task,the data set is more difficult to obtain and the labeling work is too expensive.In order to solve these problems,a convolutional neural network expression recognition method based on transfer learning is designed.This method adopts a transfer learning model and constructs a loss function based on log-Euclidean distance.The model contains two convolutional neural network channels of the source domain and the target domain.Both the source domain and target domain use the proposed network for facial expression feature extraction and share the same weight parameters.At the same time,the full connection layer is used as the domain adaptation layer to realize feature adaptation.Then,according to the constructed loss function based on the log-Euclidean distance,the distribution distance between the source domain and the target domain features is expressed,which is regarded as the correlation difference between the two domains.In the process of parameter updating,the classification loss and the log-Euclidean loss are combined for joint training to minimize the correlation discrepancy between the source and the target domain.At the end of the training,the two kinds of loss are balanced,and finally the target domain can be effectively classified.The transfer learning model solves the problem of large amount of data dependence and the same data distribution of facial expression recognition task using convolutional neural network at present.At the same time,the loss function based on log Euclidean distance can effectively solve the manifold problem of measuring the distance between two domain eigenvectors in transfer learning.Finally,the data set Raf-DB is used as the source domain,and the data set Jaffe and CK+are used as the target domain respectively.By comparing the expression recognition accuracy of ordinary convolutional neural network model and transfer learning model,it is found that the proposed transfer learning model successfully realizes the knowledge transfer across databases,and improves the facial expression classification ability of the model.