Research On Facial Expression Recognition Algorithm Based On Transfer Learning And Feature Fusion

Facial expression recognition is one of the important areas of facial attribute research.The results of facial expression recognition have been widely used in business analysis,safe driving,online education,smart medical and other fields.Therefore,the research on facial expression recognition is of great significance.However,traditional facial expression recognition algorithms need to manually design feature extraction methods,which leads to insufficient representation of features and poor model generalization capabilities.Secondly,the facial expression recognition algorithm based on deep learning mainly extracts the deep features of the image while ignoring the shallow and middle features of the image.Moreover,most of the current expression recognition data sets are small data sets,which are not enough to support the training of large-scale convolutional neural networks.In order to solve the problem of insufficient feature representation of traditional facial expression recognition algorithms and small data sets,the paper proposes a facial expression recognition algorithm based on transfer learning.According to the theory of transfer learning and related experience,the paper will migrate the Inception V3 and Res Net50 models pre-trained on the Image Net dataset to the new datasets KDEF and Ra FD,then build a classification layer to form a new network model.After determining the initial learning rate,optimizer,data preprocessing method,and model training method,modify the parameters according to the experimental results to obtain different models.Through the analysis of the experimental results,the model with the highest accuracy rate in the current data set is found.The model with the highest accuracy rate in the KDEF data set is KI21A121,with an accuracy rate of 93.77%,and the model with the highest accuracy rate in the Ra FD data set is RI21A121,with an accuracy rate of 98.98%.In order to solve the problem that the facial expression recognition algorithm based on transfer learning proposed in the paper does not have the ability to extract the shallow and middle features of the image,the paper proposes a facial expression recognition algorithm based on transfer learning and feature fusion.The paper takes the different layers of the Inception V3 or Res Net50 model pre-trained on the Image Net data set and merges them to obtain the final feature extraction layer,which is combined with the two fully connected classification layers to form a new model after fusion.After the fusion of the Inception V3 model and the Res Net50 model,the experimental results on the KDEF and Ra FD data sets show that the accuracy of the fusion model is higher than the accuracy of the model before the fusion.The fusion model KI20AFFF2 has an accuracy rate of 94.48% on the KDEF data set,and the accuracy rate of RI21AFF3 on the Ra FD data set is 99.09%.Finally,the models KI20AFFF2 and RI21AFF3 are used to identify randomly selected images in the KDEF and Ra FD data sets.The results of multiple recognition are correct,but the generalization ability of the model RI21AFF3 is better than that of KI20AFFF2.Therefore,using RI21AFF3 to recognize real faces,the results of multiple recognitions are correct.