Adversarial Synthesis Of Facial Action Units With Regional Attention

With the development of artificial intelligence,human centered facial expression recognition has been paid more and more attention in the industry and academia.Facial action units(AUs)is an important objective description of facial expression,and the detection and analysis of AU is of great significance for understanding and describing facial expression.However,due to the complexity of AU annotation and the relatively small AU expression data set,how to expand the existing AU data set to solve the label imbalance problem of AU data set has become an urgent problem in this field.Based on the method of conditional adversary network,this paper proposes a method of local attention conditional generative advantage network(LAC-GAN)to synthesize the occurrence states and intensities of many kinds of AU.Firstly,the AU attention region is calculated according to the local AU segmentation rules,and the local feature of the global face is extracted to get the corresponding AU region of interest.Then,for the provided AU tag,the conditional countermeasure generation network synthesizes the AU States and strengths of the corresponding AU local regions of interest.Finally,the attention mechanism is used to ensure the consistency between the concerned area and the non-concerned area,and to classify and distinguish the global face,so as to achieve the natural and real image quality.Based on the method adopted in this paper,it can not only realize the synthesis of a single AU,but also realize the synthesis of facial expressions with a certain degree of naturalness,which can increase the diversity and universality of AU expression data set,so as to solve the problems of insufficient AU detector training data labeling and imbalance of sample labels.In order to verify the effectiveness of the proposed method,qualitative and quantitative verification methods are used to evaluate the model.Qualitative evaluation is mainly to evaluate the quality of image generation.The quantitative evaluation is mainly to evaluate the AU quantitative index of the generated image,in order to verify that the generated image is effective for the current mainstream AU detector.