Research And Implementation Of Facial Expression Recognition Algorithm Of Cartoon Character Image

The emotional interaction between virtual characters and human beings has always been expected by people,and the premise of emotional interaction is that virtual characters can carry out emotional analysis.If virtual characters want to have the ability of emotion analysis,they must be able to recognize emotion first,so this paper is the task of emotion recognition.Virtual characters eventually need and people interact,it in the process of interaction with the help of his understanding of expression to identify people’s facial expressions,so the thesis chose cartoon facial expression data for training,real facial expression data for testing,expect to achieve cross-domain identification,for the development of facial expression recognition and realization of virtual characters and human interaction provide certain basis.Since the distribution of facial expression features of cartoon face and real face is different,this paper first converts cartoon face and real face into the same style of data set,so that cartoon face and real face use the same set of expression.Then,using transfer learning method,cross-domain emotion recognition is performed on the transformed features.On the task of image style conversion,this paper proposes a style conversion method based on unpaired data samples.The method takes generating adversarial network as the general basic network structure,and then combines hourglass network,attention mechanism and basic network to improve the new network structure.The improved network makes use of the spatial relation of the image and the weight relation of each region of the image to make the transformed image clearer and better retain the details of the original image.Experiments are carried out on JAFFE and CK+ data sets,and the results show the effectiveness of the improved network in the task of image style conversion.For emotion recognition task,this paper proposes an expression recognition method based on transfer learning combined with attention mechanism(SGE)and residual network(Res Net).In this method,a new Res-SGE basic module is constructed by combining Res Net basic module and attention mechanism module,and this module is applied in Res Net.Then,the improved Res Net network is used to build the basic network of transfer learning,and MMD distance loss function is added to measure the distance between source domain and target domain.Through experimental verification,the recognition accuracy of the proposed facial expression recognition method on the data set JAFFE and CK+ is increased by 7.55% and 10.85% respectively,indicating the effectiveness of the proposed facial expression recognition method.