Research On Facial Expression Recognition Method Based On Curriculum Learning And Measurement

As the basis of human-computer interaction,emotion understanding has attracted more and more attention in the field of service robot.As one of the most direct ways to emotion understanding,facial expression recognition still has two unsolved problems.Firstly,the generalization performance of the model trained by a large number of samples is poor,the ability of a single model to recognize seven basic expressions varies.Secondly,large sample size is required for the optimization of facial expression recognition model,at the same time,it is difficult to obtain labeled data,and image labeling is time-consuming and labor-intensive.In view of the above two problems,this paper proposes a facial expression recognition method based on course learning and measurement,and makes an in-depth study of the above two difficult problems.Aiming at the problem of poor generalization performance of facial expression recognition model,a Curriculum Learning-based facial expression recognition method is proposed,which is mainly divided into two stages:curriculum design and curriculum learning.In the stage of curriculum design,seven expressions were divided into three subsets with different complexity in a 6:2:2 ratio by using the Density-Distance unsupervised clustering method.Curriculum learning is the process of model optimization from simple to difficult.Two ways,Adding and Replace,are proposed to increase the complexity of training,and then the accuracy of subsets is proved by means of Replace.Furthermore,at the level of expression,the confusion matrix of a single model was analyzed by Euclidean distance,and the most three difficult expressions of anger,fear and sadness were obtained.The Self Selection Mechanism(SSM)was proposed to re-judge the three expressions in the test stage.Experiments have proved that the facial expression recognition method proposed in this paper achieves the optimal results of 72.11%and 98.18%respectively on FER-2013 and CK+..In addition,the cloud deployment and invocation of the facial expression recognition service is studied,and the accuracy and service time of the cloud service are tested.The test results show that the cloud service can meet the real-time requirements of the service robot.Aiming at the problem that the required sample size for the training of facial expression model is too large,and it is difficult to obtain labeled data,and image labeling is time-consuming and labor-intensive,a Prototype-Relation Network architecture is proposed to transform the image classification problem into the nearest neighbor problem.After that,the loss function of few-shot learning is improved,and the in-class distance is introduced into the loss function in the form of regularization term.The whole model is optimized in an Episode-based way to construct a 7-way K-shot training scene.Experiments on the FER-2013 dataset show that:when the data volume is 100%,80%and 60%respectively,the model accuracy is improved to different degrees after the in-class aggregation loss is introduced.When the regularization ?=0.3,the optimal recognition effect is achieved.By further exploring the relative size relationship between Ktest and Ktrain,it is proved that when Ktest>Ktrain,only 65%of the dataset can achieve the original recognition accuracy of 100%dataset,and the dataset can be reduced by 35%.Therefore,the purpose of learning with fewer samples is achieved.