Face Recognition And Facial Expression Analysis Based On Active Appearance Models

Face image analysis has been widely used in identification, security, animation driving andfilm making. As two important research aspects of face image analysis, face recognition andexpression analysis involve many fields including image processing, computer vision, virtualreality, human-computer interaction and have important theoretical and practical value. Currently,most of face image analysis systems are based on texture information and influenced byacquisition light, background and non-rigid characteristics of human faces, therefore, of featureextraction is inaccurate. It is the main development direction of face image analysis in the futureto integrate texture and geometry information for facial feature extraction..Active Appearance Models (AAM) is a method for image feature extraction proposed byCootes at the end of20th century. It uses the statistical model of combination of shape andtexture information and non-linear optimization fitting in order to extract facial featureinformation more comprehensive and accurate. This thesis applies AAM to extract the facefeature points and constructs corresponding characteristic vectors separately for face recognitionand expression analysis. The main content includes as below:1. Building AAM and selecting proper matching algorithms. Obtaining the changes in shapeand texture from a representative training set, then AAM algorithm applying PCA to give aparameterized facial appearance model, finally, using the reverse combination algorithm tocalculate the optimal matching between input image with the model and locating facial feature.2. Constructing feature vectors using the extracted features for face recognition. Firstly, takingthe feature points obtained by the AAM as reference, we established a geometric feature vectoras the critical feature to recognize face, Which vector with the scale, rotation and displacementinvariant. Then, Through the vector we propose a pre-classification in recognition library thatcould effectively reduce the number of contrasting recognition processes. Finally, byconstructing similarity function we could proceed to match image which to be tested and imagein recognition library and get final result of recognition.3. Constructing relative motion vectors using AAM features for expression analysis.According to different displacement of features achieved from different image expressions, weconstruct relative displacement vector as input features of expression recognition. Though twoways could achieve expression recognition: establish multiple SVMs to identify relativedisplacement vector as six expressions directly; establish two classifier SVMs to detect human face’s Active Unit（AU）, and judge expression through the value of AU which concluded indetected area.Experiments are made to the above algorithms and results show that using AAM for facerecognition and expression analysis can effectively avoid influence of light, background andpostures and has good feasibility and robustness. Finally, there is a conclusion and prospect ondevelopment of AAM and human facial image analysis.