Research On Facial Expression Recognition Method Based On Attention Mechanism And Double Label Learning

Facial expressions play an important role in the process of information exchange.The importance of automatic facial expression recognition in human-computer interaction is self-evident.In addition,facial expression recognition is widely used in fatigue driving detection,online education,telemedicine,game entertainment and other fields.Facial expression recognition faces many challenges in the development process.There are serious shortages of expression datasets for specific scenes,such as occlusion,illumination,label noise,and multi-expression scenes.Due to the uncertainty of the actual scene,facial expressions are easily occluded by other objects,which affects the accuracy of expression recognition.During the labeling process,the ambiguity of some expressions and the subjectivity of the annotators make the expression dataset contain a lot of label noise,which will seriously affect the training of the model.Aiming at the two problems of occlusion and label noise,this study makes the following contributions:(1)An occluded expression recognition network based on face and multi-selfattention is designed,which combines the idea of facial attention,self-attention mechanism and dual-branch learning.The network is first concentrated on the face area,and then the area is divided,and the area weight is allocated reasonably with the help of the RRB loss function.When occlusion occurs,the model assigns larger weights to important or non-occluded regions to counteract the negative effects of occlusion and improve the performance of the model.On the CK+,RAF-DB and FERPlus datasets,the model achieved 98.08%,87.74% and 88.87% accuracies,respectively,which were better than the state-of-the-art algorithms at that time.(2)I construct an expression recognition framework based on double-label learning,which is the first three-phase double-label learning network.The three phases in the training process are: warm-up phase,selection phase and relabeling phase.In the warm-up phase,all training images are used to ensure the feature representation capability of the model.In the selection phase,the reinforcement learning of the specified samples is performed.In the relabeling phase,the potential wrong labels are relabeled,and the second label is learned.The relabeling phase of the model can correct errors for samples with label noise in combination with other clean samples,thereby reducing the impact of noise and improving the robustness of the model to label noise.The model achieves the best performance on the three datasets RAF-DB,FER2013 and Affect Net.(3)In order to facilitate the testing of the model,a synthetic occlusion expression dataset is created for the occlusion problem,including two types of landmark occlusion and random occlusion.For the label noise problem,label noise datasets with multiple noise levels are created,which can be used to effectively test the performance of the model.