A Research Of Domain Adaptive Expression Recognition

In the facial expression recognition problem,extracting facial expression features is a very important step,and the quality of the features directly determines the recognition effect.Although the area of the face area is not large,it does contain a lot of face attributes.The feature of face attributes that is not related to the expression during the feature extraction process will have a negative impact on the classification effect and increase the burden on the classifier.At the same time,the environment in which the sample is located will also affect the process of feature extraction and affect the recognition effect.The ultimate goal of the research on facial expression recognition is to apply this technology to daily life,but the facial expression recognition algorithm needs to meet the premise of independent and identical distribution of training samples and test samples to achieve better results.This condition is very difficult to meet in practical applications.Therefore,this article is doing the following work on these issues:1.In the process of facial feature extraction,in order to obtain the facial expressionrelated features,other irrelevant facial attribute features are removed.This paper introduces the spatial attention mechanism of images to feature extraction,and designs a spatial attention extraction network.By minimizing the loss function of the spatial attention extraction network proposed in this paper,the spatial attention extraction network can focus on the face organ regions with strong expressions.The features of those regions are enhanced and the features of ther regions are weakened to achieve the purpose of enhancing expression attribute characteristics and weakening other facial attribute characteristics.At the same time,according to the knowledge of facial anatomy,the facial muscle trajectories of the same type of expression are similar,and the facial muscle trajectories of different expressions are different.The attention area in this article should also meet this condition.The expressions are similar,and different expressions are different,so the experiments in this paper visually verify it,and prove that the attention area obtained by this part of the network is effective.At the same time,according to the knowledge of facial anatomy,the facial muscle trajectories of the same type of expression are similar,and the facial muscle trajectories of different expressions are different.The area of interest in this article should also meet this condition.The expressions are similar,and different expressions are different,so the experiments in this paper visualize the attention results and prove the effectiveness of the algorithm.2.In order to extract the features that only include facial expression features,and to filter out irrelevant facial attribute features,this article also applies the knowledge of domain adaptive learning to expression recognition.The premise of domain adaptive algorithm migration is to extract public features The common feature between the expression database data is the expression feature,and the domain adaptive algorithm can filter out the face attribute features that are not related to the expression to a certain extent.This paper proposes an unsupervised domain adaptive algorithm,which improves the traditional decision-making process without considering the decision boundary of the classification problem.At the same time,the unsupervised algorithm does not require expression tags on the training data of the target domain during the migration process.In practical applications,face samples without expression tags are very easy to obtain,and can obtain more training samples,making training The sample distribution is closer to the real-world sample distribution,improving the generalization ability of the trained model.At the same time,in this paper,experiments on cross-datasets are performed on commonly used facial expression databases to verify the algorithm.The algorithm in this paper has achieved the world's highest recognition accuracy on cross-dataset experiments.Experiments demonstrate that our algorithm exceed state of the art results for crossdataset experiments.