Design And Implementation Of A Deep Neural Network For AU Detection System

Using deep learning to solve real-life image problems is becoming an important research topic in the field of computer vision,and the research of computer vision methods for facial expressions has become one of the important directions.Compared with expressions in the traditional sense,micro-expressions have more important significance in some fields,such as medicine,question-answer,and trial.It is of great value to use deep learning methods to achieve automatic detection of micro-expressions.In this thesis,the micro-expression is defined by Facial Action Coding System(FACS),which is called Action Unit(AU).This thesis focuses on how to achieve automatic facial AU detection using effective deep learning algorithms.Currently,the implementation of facial AU detection using deep learning methods faces some bottlenecks.This thesis focuses on the region problem,person-specific feature overfitting problem,label noise and data quantity problems of AU detection.To address these problems,this thesis proposes:(1)A Multiple Region Perception module and a Pixel-Interest learning strategy for capturing fine-grained feature of AU and region relationship.This relationship is self-learning and better than pre-defined;(2)An Anti Person-Specific learning strategy to eliminate person-specific features,which can effectively suppress the overfitting on individual personal features;(3)A Semi-Supervised learning strategy with Discrete Feedback(DF),which is based on a pseudo label learning pipeline with additional discrete feedback.The discrete feedback can effectively guide the teacher network to converge and suppress the noise in the labels.By combining the three strategies proposed above,this thesis implements a complete pipeline of deep learning for face AU detection and improves the performance of the model.Detailed ablation experiments and comparison experiments are given to verify the contribution of each strategy and prove the superiority of the method.The algorithm proposed by this thesis has reached the SOTA(State Of The Art).