Research On Cartoon Face And Real Image Emotion Recognition

Augmented reality(AR)has a wide range of application scenarios and markets.It integrates information from the virtual world and the real world to give humans an experience that transcends reality.With the prosperous development of the animation and game industries,the demand for AR interaction between virtual characters and people has been dramatically increased.The AR interaction technology of virtual characters between people is highly anticipated by people,and it can be applied to many fields such as animation,games,psychotherapy,and so on.At present,the interactive technology has limited abilities of "listening","speaking","reading" and "watching",but it cannot realize the emotion of the interactor.People's emotional state greatly affects their behavior,so this paper hopes to introduce emotion analysis into this interactive technology to make it more intelligent and humanized.However,due to the limitations of the conditions,this paper can only regard the interactive interface of virtual characters and real humans as the combination of anime characters and real image scenes.According to this,this paper researches the emotional recognition of anime character faces and real image scenes,which can provide a solid technical foundation for the sentiment analysis of the AR interaction between virtual characters and humans.In this regard,the main research contents of this paper are as follows:(1)A cartoon-to-real face translation algorithm with Generation Adversarial Networks is proposed.After cartoon faces are transferred to the real ones by this algorithm,the real facial expression recognition method is used to analyze the emotion of cartoon faces with the translated faces.Because it is difficult to collect enough facial data of anime characters with emotional labels,and the facial expression recognition model of anime characters cannot be directly trained based on the collected facial data of anime characters without labels,this paper proposes an unsupervised facial expression recognition scheme for anime characters.Generative adversarial network is used to convert the anime face to the corresponding real face,and then the emotion recognition of the converted face is performed using the existing real face expression recognition model.In order to ensure that the converted face has intact and clear facial features,this paper uses global and local discriminators to capture the global features and three local features of the face,respectively.Besides,this paper uses the content network to preserve facial features and identities between cartoon and real faces.Experiments show that compared with the general methods of image translation in recent years,the proposed method can transform anime faces into more realistic and relevant faces.Finally,the experiment also shows that emotion recognition scheme using the converted faces is more effective than the emotion recognition directly using the unlabeled cartoon faces.(2)An image scene emotion recognition algorithm with Massage Passing Neural Network is proposed.Current methods on image scene emotion recognition use attention models,saliency detection,or region proposals to point out the most influential regions of emotions.However,different instances in the image can have different effects on emotions.In order to obtain a good emotional expression,it is not enough to just focus on the entire image or a single local area.Therefore,in this paper,through the emotional interaction of multiple instances in the image,a more expressive image emotion comprehensive expression is obtained.More specifically,this paper introduces the Massage Passing Neural Network into the image scene emotion recognition task and proposes an image emotion recognition model that can realize multi-instance emotion interaction.Experiments show that the proposed method is superior to the current state-of-the-art methods in image scene emotion classification tasks.