Research On Recognition Algorithm Of Facial Action Coding System

In addition to language information,the information carried by human facial expression has always been one of the researchers' concerns.With the rapid development of computer vision and machine learning,Facial Action Coding System proposed by micro-expression related scholars is a system that encodes different motion categories of each facial muscle based on small movements of human facial appearance.How to automatically and accurately identify each AU(AU,Action Unit)label in the expression image,so as to identify a variety of facial expressions in the image is a challenging task.Therefore,this paper combines feature detection and deep learning to propose a new Facial Action Coding System recognition algorithm,focusing on issues,such as the expansion of facial expression dataset with Action Units label,extraction and visualization of features in facial expression images,and application of deep convolution network model in recognition of Facial Action Coding System.The main research contents are as follows:Firstly,because the research on Facial Action Coding System relies on large-scale facial expression datasets with precise labels,but at the present stage,artificial annotation is highly professional,difficult and complex.Therefore,this paper proposes a method based on generating adversarial networks to expansion the facial expression dataset.We design a new loss function,use the expression images and the corresponding AU labels to train the model.According to the target Action Unit label,the model can automatically generate the facial expression image corresponding to the target label.We achieved that the generation of real facial expression images is controlled by multiple action units labels,which is the purpose of expanding the expression dataset.Secondly,since the fundamental purpose of Action Units recognition is to identify the motion category of each facial muscle unit,this paper uses the pore-scale feature point detection algorithm and the Grid-based Motion Statistics to match the feature points of two consecutive video images.The magnitude and direction information of the relative displacement of the feature points are transformed into the vector field.We use the Munsell color system to effectively and intuitively visualize the direction and amplitude information of each feature point motion generated by the movement of Action Units in each facial image.Finally,in order to improve the performance of the existing Facial Action Coding System recognition algorithm,this paper combines the information of visualizing the displacement of pore-scale feature points,applies classic deep learning network models,and selects the best module design solution through experimental comparison.On the common BP4 D facial expression dataset,compared with a variety of existing recognition algorithms,we obtained the best comprehensive recognition effect of Facial Action Coding System.Among them,the average accuracy rate of the Action Units recognition of the algorithm proposed in this paper obtained a better 75.88% among the various recognition algorithms,at the same time,the relative highest average F1 score was obtained as 63.46%.