Research On Expression Recognition Method Based On Transfer Learning And Attention Mechanism

Facial expression recognition needs to train a large amount of data through convolutional neural network,but the number of facial expression samples in many public data sets is small,which easily leads to overfitting of the model on the target data set.Facial expression recognition is mostly applied in open scenes,where images input from the outside may be mixed with irrelevant images,and neural networks may be disturbed by irrelevant images,resulting in a decrease in accuracy.Therefore,this paper adopts the expression recognition method based on transfer learning and attention mechanism.The research work of this paper is as follows:(1)For many public data sets,the number of facial expression samples is small,and the expression information is single,which is not conducive to the research of facial expression recognition.This article collected facial expression images manually and created a facial expression dataset for experiments.(2)A facial expression recognition method based on transfer learning and hierarchical attention mechanism is proposed.Firstly,the ResNet network is improved by incorporating the residual block structure of the Res2Net model and incorporating the CBAM attention mechanism after the residual block.Pre train the improved model on ImageNet,and then transfer it to a facial expression dataset with relatively few samples for training.Experiments were conducted on the improved ResNet network model using self built datasets and public datasets CK+,respectively.(3)An open set facial expression recognition method based on attention mechanism was proposed.Combine ResNet network integrated with CBAM attention mechanism with OpenMax open set expression recognition method.A feature extraction network incorporating CBAM attention mechanism was proposed to recognize facial expression features in open scenes.The experimental results show that the method proposed in this paper can make the model focus on finer features and improve the recognition accuracy of known class recognition under the open set setting.