A Study And Implementation Of Face Alignment And Facial Expression Recognition Algorithm

There are a lot of mature techniques in Computer Vision field,e.g.face recognition.There have been some commercial face verification systems in the industry.Facial Recognition is similar with face recognition while they are different in some aspects as well.They are both face problems and are sensitive to illumination and pose.The difference is that face recognition issue is to distinguish identities regardless of the expression while facial expression recognition is to recognize the identical expression on different faces.A lot of features for identity recognition have been used for facial expression recognition.Facial Expression Recognition has been a fundamental project in CV area.However,there are not many fabulous works.Although Deep Learning has become rather popular in a wide range of field of AI,it is rarely used for recognizing facial expression.The biggest obstacle is the shortage of training samples.This paper mainly focuses on expression related algorithms,including facial landmarks detection,face pose estimation,expression feature design and so on.We have built a expression recognition system which can detect facial landmarks and estimate head pose in real-time and recognize macro facial expression with high accuracy.(1)Landmark Detection:Facial landmarks are those key points of the corners and edges of facial features.Detecting facial landmarks is a preprocess for expression recognition.Most hand-crafted features for expression recognition are based on the coordinates of landmarks.We will study the current state-of-art face alignment algorithms and re-implement a fast one which is called 3000FPS.Our facial landmark detection system can detect facial landmarks in real-time.(2)Face pose estimation:After the landmarks are detected,a lot of work can be done with the landmarks.Face pose estimation is one the application.We will make use of the coordinates of 2-D facial landmarks to estimate the Eular Angles of a given face roughly.(3)Expression Recognition:Facial expression issue is much like other pattern recognition problems,which have two main steps:feature extraction and classifier training.We designed two efficient expression features in this paper:geometric shape increment feature and texture feature.In the geometric shape increment feature,the mean shape of Neutral faces is calculated first.The difference between a face sample and the Neutral pattern consists of the feature.The increment depicts the moving trace of each landmark.which includes facial expression information.In the texture feature,a gradient histogram of the face grids is generated.This feature can describe the texture information of the whole face.Geometric features are robust to illumination while texture features are robust to face pose.We also propose a method to fuse the generative outputs of different models,which has improved the recognition performance greatly.We will illustrate the efficiency of the landmark detection and expression recognition algorithm with experiments in the latter chapter.