Study On Facial Expression Recognition Of Students Based On Attention Mechanism Convolutional Neural Network

At present,due to the impact of the COVID-19 epidemic,the combination of online and offline classes has become a new trend in education.The recognition of students' facial expressions can be used as a means to assist teaching,helping to improve teaching quality from the side,education and teaching will develop towards the direction of truly meaningful and valuable teaching.With the rapid development of computer vision,facial expression recognition as a key research topic has been gradually extended to various fields.Traditional deep learning methods have improved the recognition rate of facial expressions to some extent,but the difference of facial expressions is not very obvious,leading to poor recognition accuracy of some facial expressions.Therefore,their application is very narrow in the educational environment of "Internet + education".1)Firstly,an attention mechanism convolutional neural network for A-IM-Conv LSTM recognition framework is proposed.The video frame and optical flow images were used as double inputs to the A-IM-CONVLSTM network.The additional explanatory factors were extracted by improving the traditional convolutional long and short time memory mechanism network.At the same time,multi-feature learning is carried out on video frames and optical flow images,and attention mechanism module is added to infer the non-occlusion degree of video images,and weighted representation of feature vectors is carried out.Attention deviation is carried out on images with small occlusion degree to form decisions.Finally,all weighted feature vectors are fused and sent into the full connection layer.2)A new discriminant distribution unknowable loss(DDU)is proposed to optimize the classification of extreme expression in the case of imbalanced categories.The unbalanced distribution of some expression categories leads to confusion of expression categories in recognition.The proposed DDU loss function forces inter-class separation of the deep features of most categories and few categories.In order to make the mapping from embedded space to label space more effective,a well-separated depth feature cluster is generated in the embedded space,and the relevant information in the embedded space is selectively clustered and separated between classes.Finally,experiments are carried out on the public FER dataset.Compared with several advanced FER methods,the proposed method can enhance the ability of feature recognition and improve the accuracy of facial expression recognition.