Research On Approaches Of Emotion Recognition In Natural Interactive Scene

Emotion recognition in the wild is of great significance for affective computing and human-computer interaction systems,but there are still some problems in technical research,such as large differences in face quality and complex and diverse context information.In order to solve the above problems,this thesis focuses on emotion recognition tasks for natural interaction,and studies key technologies such as emotion recognition based on facial and pose feature fusion,emotion recognition based on attention and graph convolutional networks.To solve the problem of uneven face quality in natural interaction scenarios,we design an emotion recognition system based on facial quality perception and gesture recognition.Specifically,for facial features,we first use face detection methods to detect faces,and then use the face image quality assessment module to obtain face quality scores.For higher quality face images,we perform action unit detection,and according to rule of mapping from AUs to emotions,we design a three-layer Bayesian network to map detected action units to emotion categories.For low-quality face images,we directly extract the entire face features using CNN.In addition,we use the body posture features to make up for the lack of face information.Concretely,we perform pose estimation at first,then perform feature extraction based on these detected pose estimation pictures,finally combine face and pose features to predict the emotion category together.In order to solve the complex and diverse context information problem,we build a key emotional elements detection and reasoning network based on attention and reasoning mechanisms.First,we send the pictures into the convolutional neural network to initially extract semantic features.In addition,we construct an attention local feature descriptor to learn the weights of the extracted semantic features.The weight indicates the contribution value of this local feature to the emotion recognition task.After obtaining the features of the key emotional regions,the graph convolutional neural network is used to learn the emotional relationship between different regions,and finally infer the emotional state.The above algorithms have achieved higher accuracy in emotion recognition than other methods on currently popular emotion recognition datasets.