Research On Classroom Expression Recognition Based On Deep Learning

With the rapid development of modern artificial intelligence technology and the deepening of education reform,the education field has ushered in a new era of intelligent education.In traditional classes,if teachers want to acquire students’ emotional state in class in real time,they mainly ask questions on the spot and observe by naked eye,which requires extra energy and is highly subjective.Therefore,it is of great value and significance to use intelligent recognition technology to automatically acquire students’ emotional state in class by facial expression recognition.In order to further the study of facial expression recognition in class,the following work is done:(1)By referring to the research results of students’ emotional states in distance learning and offline classes and combining with the characteristics of students’ learning environment,a new classification of classroom expressions was proposed: listening,understanding,thinking,wandering,yawning.The classroom expression data set is constructed and a variety of data enhancement methods are selected to expand the data set.(2)This paper proposes an attentional mechanism based classroom student expression recognition method,and the accuracy rate is 95.48% after experiments,which effectively improves the network performance.Firstly,the attention mechanism is introduced into the backbone network to make the network focus on the most prominent features of the input image.Then,the activation function is modified,and the linear operation is used to enhance the feature performance and reduce the number of parameters and computation.Finally,the Dropout regularization is added after the pool layer to prevent overfitting of the model.(3)Aiming at the problems of insufficient data quantity and uneven types in the current data set,a data enhancement method based on generative adduction-network is proposed.Experiments show that the overall quality of the generated images is improved,and the accuracy of expression recognition by using the generated data is improved to 97.58%.Firstly,Star GAN network with high conversion efficiency is selected as the basic model.Secondly,the Multi Res UNet model is introduced as a network generator to make the edge feature extraction more accurate.Finally,the improved reconstruction loss function helps the network to distinguish significant defects and tiny image displacement in key areas.