Research On Disequilibrium Processing Of Facial Action Unit Recognition

Emotion is a subjective experience of human beings,and its external expression is mainly through facial movements.Traditional research on facial movement recognition mainly focuses on identifying six kinds of racial and culturally unrelated expressions,reflecting the basic emotions of human beings.However,it is clear that the six expressions do not fully describe the entire facial movements and emotions of the person.The facial action coding system is recognized as the best mechanism for describing facial movements,and various expressions can be broken down into a series of basic combinations of facial muscle movements for subsequent coding.This subject is derived from the National Natural Science Foundation's general project "Recognition of Natural Human Facial Action Units in an Open Environment".This topic aims to take the recognition of natural human action units as the research object,and to start from the key factors that affect the performance of the recognition model and carry out related research on the characteristics of the problems for the imbalance of facial action unit categories.Unbalanced facial action unit category refers to the unbalanced distribution of action unit categories,resulting in the phenomenon that model recognition is biased to the majority of samples.Under the selected action unit data set CK+,the conditional generation is used to improve the class imbalance of the action unit against the network,and a support vector machine is designed to perform two-class classification to verify the improvement effect of the conditional generation against the network generated samples on the class imbalance.Based on clear research background and research significance,this paper investigates and analyzes the research status of action unit category unbalanced processing methods at home and abroad,and finally completes the design and verification of the overall algorithm.In this paper,we study the processing method in the case where the category imbalance phenomenon causes the model identification to be biased to the majority sample,improve the training data in terms of quantity and quality,and reduce the impact of category imbalance.The algorithm principle of generative adversarial networks and conditional generative adversarial networks is studied,and the model parameters and algorithm flow are designed.Complete the selection and construction of generated network and discriminated network parameters in the model.Design the model training process by setting up the environment on the Didi cloud GPU server.After the model training,the image quality of the generated samples is good and the loss function converges.The design completes the process design of the binary classification algorithm based on support vector machine.After processing the samples,the histogram of the direction gradient is used for feature extraction,and then the extracted image features are subjected to principal component analysis to reduce the dimension.The classifier parameter design is carried out by constructing a visual bag of words model and selecting a classifier kernel function.Set up different sample groups: select classified training and test sample data,conduct binary classification model training,and use F1 score to analyze the effect of binary classification,thus proving the effectiveness of the generated samples to improve the imbalance of categories.Through experiments,it can be concluded that the conditional generative adversarial network has an obvious improvement effect on the unbalanced action unit category,and has reached the research goal of the subject..