Research On Image Emotion Recognition And Its Application In Teaching Robots

With the development of artificial intelligence,image emotion recognition has become a hot research problem in the field of human-robot interaction.Image emotion recognition includes expression recognition and behavior recognition,which is one of the key technologies in the field of educational robots,and currently there are problems such as poor expression recognition ability and low behavior recognition rate.And teaching robots as one of the main objects of future interactive teaching mode,emotion recognition technology and its application on teaching robots have important research value.The main research contents of the paper are:First,an expression recognition method with residual attention mechanism and LBP feature fusion is proposed to address the problem that expression recognition in a natural reality environment is affected by head pose variation,uneven lighting,occlusion and individual variability of facial expressions,and category imbalance between different datasets,which leads to low accuracy of expression recognition.The method solves the problems of imbalance between cross-library datasets,low recognition rate,and better generalization ability and robustness in the laboratory and natural reality environments.Compared with the baseline Res Net18,the recognition accuracy is improved by 7.37%,5.94%,2.69%,and 5.28% on the CK+,FER2013,FERPUS,and RAF-DB datasets,respectively.The approach of the augmented residual attention model incorporating LBP features and attention mechanism was verified to be feasible by conducting ablation experiments and expression heat map visualization experiments.Secondly,an improved lightweight expression recognition method is proposed for the application of expression recognition models in mobile or embedded devices in a teaching environment,which faces problems such as a large number of model parameters and slow computation speed,resulting in reduced recognition accuracy.The model of this method has the characteristics of lightweight and portability.Experiments were conducted on the RAF-DB and FER+7 expression datasets,and the recognition accuracy was 87.35% and88.48%,respectively.By comparing with the classical model,the computational speed and the recognition accuracy are improved while maintaining a smaller number of parameters.Then,an improved behavior recognition method is proposed as an important complement to emotion expression in a teaching environment can further improve the emotion recognition accuracy.To more accurately obtain the emotional states and intentions of people in the teaching environment,a self-built behavioral dataset was constructed,a Mobile Net network model was used for skeletal point feature extraction,and different classifiers SVM,KNN,and MLP were used to compare the recognition results,and the behavioral recognition accuracies were 93.61%,92.84%,and 95.75%,respectively,compared with the original model VGG19,which improved compared with the original model VGG19 by 1.24%,2.79%,and 3.27%,respectively.The experimental results show that the method has better feature extraction ability and classification performance in the behavioral recognition task in a teaching environment.Finally,for the lack of emotion recognition capability of the teaching robot,an emotion recognition system is designed and the constructed lightweight model is integrated into the NAO robot to achieve the teaching human-robot interaction goal of the NAO robot,and the proposed expression recognition method and expression recognition function are verified,and the results when the modal fusion of expression and behavior are used to determine the human emotional state.